Speech-to-speech translation (S2ST) is key to breaking down language barriers between people all over the world. Automatic S2ST systems are typically composed of a cascade of speech recognition, machine translation, and speech synthesis subsystems. However, such cascade systems may suffer from longer latency, loss of information (especially paralinguistic and non-linguistic information), and compounding errors between subsystems.
In 2019, we introduced Translatotron, the first ever model that was able to directly translate speech between two languages. This direct S2ST model was able to be efficiently trained end-to-end and also had the unique capability of retaining the source speaker’s voice (which is non-linguistic information) in the translated speech. However, despite its ability to produce natural sounding translated speech in high fidelity, it still underperformed compared to a strong baseline cascade S2ST system (e.g., composed of a direct speech-to-text translation model [1, 2] followed by a Tacotron 2 TTS model).
In “Translatotron 2: Robust direct speech-to-speech translation”, we describe an improved version of Translatotron that significantly improves performance while also applying a new method for transferring the source speakers’ voices to the translated speech. The revised approach to voice transference is successful even when the input speech contains multiple speakers speaking in turns while also reducing the potential for misuse and better aligning with our AI Principles. Experiments on three different corpora consistently showed that Translatotron 2 outperforms the original Translatotron by a large margin on translation quality, speech naturalness, and speech robustness.
Translatotron 2 Translatotron 2 is composed of four major components: a speech encoder, a target phoneme decoder, a target speech synthesizer, and an attention module that connects them together. The combination of the encoder, the attention module, and the decoder is similar to a typical direct speech-to-text translation (ST) model. The synthesizer is conditioned on the output from both the decoder and the attention.
There are three novel changes between Translatotron and Translatotron 2 that are key factors in improving the performance:
More Powerful and Responsible Voice Retention The original Translatotron was able to retain the source speaker's voice in the translated speech, by conditioning its decoder on a speaker embedding generated from a separately trained speaker encoder. However, this approach also enabled it to generate the translated speech in a different speaker's voice if a clip of the target speaker's recording were used as the reference audio to the speaker encoder, or if the embedding of the target speaker were directly available. While this capability was powerful, it had the potential to be misused to spoof audio with arbitrary content, which posed a concern for production deployment.
To address this, we designed Translatotron 2 to use only a single speech encoder, which is responsible for both linguistic understanding and voice capture. In this way, the trained models cannot be directed to reproduce non-source voices. This approach can also be applied to the original Translatotron.
To retain speakers' voices across translation, researchers generally prefer to train S2ST models on parallel utterances with the same speaker's voice on both sides. Such a dataset with human recordings on both sides is extremely difficult to collect, because it requires a large number of fluent bilingual speakers. To avoid this difficulty, we use a modified version of PnG NAT, a TTS model that is capable of cross-lingual voice transferring to synthesize such training targets. Our modified PnG NAT model incorporates a separately trained speaker encoder in the same way as in our previous TTS work — the same strategy used for the original Translatotron — so that it is capable of zero-shot voice transference.
Following are examples of direct speech-to-speech translation from Translatotron 2 in which the source speaker’s voice is retained:
To enable S2ST models to retain each speaker’s voice in the translated speech when the input speech contains multiple speakers speaking in turns, we propose a simple concatenation-based data augmentation technique, called ConcatAug. This method augments the training data on the fly by randomly sampling pairs of training examples and concatenating the source speech, the target speech, and the target phoneme sequences into new training examples. The resulting samples contain two speakers’ voices in both the source and the target speech, which enables the model to learn on examples with speaker turns. Following are audio samples from Translatotron 2 with speaker turns:
More audio samples are available here.
Performance Translatotron 2 outperforms the original Translatotron by large margins in every aspect we measured: higher translation quality (measured by BLEU, where higher is better), speech naturalness (measured by MOS, higher is better), and speech robustness (measured by UDR, lower is better). It particularly excelled on the more difficult Fisher corpus. The performance of Translatotron 2 on translation quality and speech quality approaches that of a strong baseline cascade system, and is better than the cascade baseline on speech robustness.
Multilingual Speech-to-Speech Translation Besides Spanish-to-English S2ST, we also evaluated the performance of Translatotron 2 on a multilingual set-up in which the model took speech input from four different languages and translated them into English. The language of the input speech was not provided, which forced the model to detect the language by itself.
On this task, Translatotron 2 again outperformed the original Translatotron by a large margin. Although the results are not directly comparable between S2ST and ST, the close numbers suggest that the translation quality from Translatotron 2 is comparable to a baseline speech-to-text translation model, These results indicate that Translatotron 2 is also highly effective on multilingual S2ST.
Acknowledgments The direct contributors to this work include Ye Jia, Michelle Tadmor Ramanovich, Tal Remez, Roi Pomerantz. We also thank Chung-Cheng Chiu, Quan Wang, Heiga Zen, Ron J. Weiss, Wolfgang Macherey, Yu Zhang, Yonghui Wu, Hadar Shemtov, Ruoming Pang, Nadav Bar, Hen Fitoussi, Benny Schlesinger, Michael Hassid for helpful discussions and support.
When a person navigates around an unfamiliar building, they take advantage of many visual, spatial and semantic cues to help them efficiently reach their goal. For example, even in an unfamiliar house, if they see a dining area, they can make intelligent predictions about the likely location of the kitchen and lounge areas, and therefore the expected location of common household objects. For robotic agents, taking advantage of semantic cues and statistical regularities in novel buildings is challenging. A typical approach is to implicitly learn what these cues are, and how to use them for navigation tasks, in an end-to-end manner via model-free reinforcement learning. However, navigation cues learned in this way are expensive to learn, hard to inspect, and difficult to re-use in another agent without learning again from scratch.
An appealing alternative for robotic navigation and planning agents is to use a world model to encapsulate rich and meaningful information about their surroundings, which enables an agent to make specific predictions about actionable outcomes within their environment. Such models have seen widespread interest in robotics, simulation, and reinforcement learning with impressive results, including finding the first known solution for a simulated 2D car racing task, and achieving human-level performance in Atari games. However, game environments are still relatively simple compared to the complexity and diversity of real-world environments.
In “Pathdreamer: A World Model for Indoor Navigation”, published at ICCV 2021, we present a world model that generates high-resolution 360º visual observations of areas of a building unseen by an agent, using only limited seed observations and a proposed navigation trajectory. As illustrated in the video below, the Pathdreamer model can synthesize an immersive scene from a single viewpoint, predicting what an agent might see if it moved to a new viewpoint or even a completely unseen area, such as around a corner. Beyond potential applications in video editing and bringing photos to life, solving this task promises to codify knowledge about human environments to benefit robotic agents navigating in the real world. For example, a robot tasked with finding a particular room or object in an unfamiliar building could perform simulations using the world model to identify likely locations before physically searching anywhere. World models such as Pathdreamer can also be used to increase the amount of training data for agents, by training agents in the model.
How Does Pathdreamer Work? Pathdreamer takes as input a sequence of one or more previous observations, and generates predictions for a trajectory of future locations, which may be provided up front or iteratively by the agent interacting with the returned observations. Both inputs and predictions consist of RGB, semantic segmentation, and depth images. Internally, Pathdreamer uses a 3D point cloud to represent surfaces in the environment. Points in the cloud are labelled with both their RGB color value and their semantic segmentation class, such as wall, chair or table.
To predict visual observations in a new location, the point cloud is first re-projected into 2D at the new location to provide ‘guidance’ images, from which Pathdreamer generates realistic high-resolution RGB, semantic segmentation and depth. As the model ‘moves’, new observations (either real or predicted) are accumulated in the point cloud. One advantage of using a point cloud for memory is temporal consistency — revisited regions are rendered in a consistent manner to previous observations.
To convert guidance images into plausible, realistic outputs Pathdreamer operates in two stages: the first stage, the structure generator, creates segmentation and depth images, and the second stage, the image generator, renders these into RGB outputs. Conceptually, the first stage provides a plausible high-level semantic representation of the scene, and the second stage renders this into a realistic color image. Both stages are based on convolutional neural networks.
Diverse Generation Results In regions of high uncertainty, such as an area predicted to be around a corner or in an unseen room, many different scenes are possible. Incorporating ideas from stochastic video generation, the structure generator in Pathdreamer is conditioned on a noise variable, which represents the stochastic information about the next location that is not captured in the guidance images. By sampling multiple noise variables, Pathdreamer can synthesize diverse scenes, allowing an agent to sample multiple plausible outcomes for a given trajectory. These diverse outputs are reflected not only in the first stage outputs (semantic segmentation and depth images), but in the generated RGB images as well.
Pathdreamer is trained with images and 3D environment reconstructions from Matterport3D, and is capable of synthesizing realistic images as well as continuous video sequences. Because the output imagery is high-resolution and 360º, it can be readily converted for use by existing navigation agents for any camera field of view. For more details and to try out Pathdreamer yourself, we recommend taking a look at our open source code.
Application to Visual Navigation Tasks As a visual world model, Pathdreamer shows strong potential to improve performance on downstream tasks. To demonstrate this, we apply Pathdreamer to the task of Vision-and-Language Navigation (VLN), in which an embodied agent must follow a natural language instruction to navigate to a location in a realistic 3D environment. Using the Room-to-Room (R2R) dataset, we conduct an experiment in which an instruction-following agent plans ahead by simulating many possible navigable trajectory through the environment, ranking each against the navigation instructions, and choosing the best ranked trajectory to execute. Three settings are considered. In the Ground-Truth setting, the agent plans by interacting with the actual environment, i.e. by moving. In the Baseline setting, the agent plans ahead without moving by interacting with a navigation graph that encodes the navigable routes within the building, but does not provide any visual observations. In the Pathdreamer setting, the agent plans ahead without moving by interacting with the navigation graph and also receives corresponding visual observations generated by Pathdreamer.
When planning ahead for three steps (approximately 6m), in the Pathdreamer setting the VLN agent achieves a navigation success rate of 50.4%, significantly higher than the 40.6% success rate in the Baseline setting without Pathdreamer. This suggests that Pathdreamer encodes useful and accessible visual, spatial and semantic knowledge about real-world indoor environments. As an upper bound illustrating the performance of a perfect world model, under the Ground-Truth setting (planning by moving) the agent’s success rate is 59%, although we note that this setting requires the agent to expend significant time and resources to physically explore many trajectories, which would likely be prohibitively costly in a real-world setting.
Conclusions and Future Work These results showcase the promise of using world models such as Pathdreamer for complicated embodied navigation tasks. We hope that Pathdreamer will help unlock model-based approaches to challenging embodied navigation tasks such as navigating to specified objects and VLN.
Applying Pathdreamer to other embodied navigation tasks such as Object-Nav, continuous VLN, and street-level navigation are natural directions for future work. We also envision further research on improved architecture and modeling directions for the Pathdreamer model, as well as testing it on more diverse datasets, including but not limited to outdoor environments. To explore Pathdreamer in more detail, please visit our GitHub repository.
Acknowledgements This project is a collaboration with Jason Baldridge, Honglak Lee, and Yinfei Yang. We thank Austin Waters, Noah Snavely, Suhani Vora, Harsh Agrawal, David Ha, and others who provided feedback throughout the project. We are also grateful for general support from Google Research teams. Finally, we thank Tom Small for creating the animation in the third figure.
Multimodal visio-linguistic models rely on rich datasets in order to model the relationship between images and text. Traditionally, these datasets have been created by either manually captioning images, or crawling the web and extracting the alt-text as the caption. While the former approach tends to result in higher quality data, the intensive manual annotation process limits the amount of data that can be created. On the other hand, the automated extraction approach can lead to bigger datasets, but these require either heuristics and careful filtering to ensure data quality or scaling-up models to achieve strong performance. An additional shortcoming of existing datasets is the dearth of coverage in non-English languages. This naturally led us to ask: Can one overcome these limitations and create a high-quality, large-sized, multilingual dataset with a variety of content?
Today we introduce the Wikipedia-Based Image Text (WIT) Dataset, a large multimodal dataset, created by extracting multiple different text selections associated with an image from Wikipedia articles and Wikimedia image links. This was accompanied by rigorous filtering to only retain high quality image-text sets. As detailed in “WIT: Wikipedia-based Image Text Dataset for Multimodal Multilingual Machine Learning”, presented at SIGIR ‘21, this resulted in a curated set of 37.5 million entity-rich image-text examples with 11.5 million unique images across 108 languages. The WIT dataset is available for download and use under the Creative Commons license. We are also excited to announce that we are hosting a competition with the WIT dataset in Kaggle in collaboration with Wikimedia Research and other external collaborators.
The unique advantages of the WIT dataset are:
Generating the Dataset The main goal of WIT was to create a large dataset without sacrificing on quality or coverage of concepts. Thus, we started by leveraging the largest online encyclopedia available today: Wikipedia.
For an example of the depth of information available, consider the Wikipedia page for Half Dome (Yosemite National Park, CA). As shown below, the article has numerous interesting text captions and relevant contextual information for the image, such as the page title, main page description, and other contextual information and metadata.
We started by selecting Wikipedia pages that have images, then extracted various image-text associations and surrounding contexts. To further refine the data, we performed a rigorous filtering process to ensure data quality. This included text-based filtering to ensure caption availability, length and quality (e.g., by removing generic default filler text); image-based filtering to ensure each image is a certain size with permissible licensing; and finally, image-and-text-entity–based filtering to ensure suitability for research (e.g., excluding those classified as hate speech). We further randomly sampled image-caption sets for evaluation by human editors, who overwhelmingly agreed that 98% of the samples had good image-caption alignment.
Highly Multilingual With data in 108 languages, WIT is the first large-scale, multilingual, multimodal dataset.
The First Contextual Image-Text Dataset Most multimodal datasets only offer a single text caption (or multiple versions of a similar caption) for the given image. WIT is the first dataset to provide contextual information, which can help researchers model the effect of context on image captions as well as the choice of images.
In particular, key textual fields of WIT that may be useful for research include:
WIT has broad coverage across these different fields, as shown below.
A High-Quality Training Set and a Challenging Evaluation Benchmark The broad coverage of diverse concepts in Wikipedia means that the WIT evaluation sets serve as a challenging benchmark, even for state-of-the-art models. We found that for image-text retrieval, the mean recall scores for traditional datasets were in the 80s, whereas for the WIT test set, it was in the 40s for well-resourced languages and in the 30s for the under-resourced languages. We hope this in turn can help researchers to build stronger, more robust models.
WIT Dataset and Competition with Wikimedia and Kaggle Additionally, we are happy to announce that we are partnering with Wikimedia Research and a few external collaborators to organize a competition with the WIT test set. We are hosting this competition in Kaggle. The competition is an image-text retrieval task. Given a set of images and text captions, the task is to retrieve the appropriate caption(s) for each image.
To enable research in this area, Wikipedia has kindly made available images at 300-pixel resolution and a Resnet-50–based image embeddings for most of the training and the test dataset. Kaggle will be hosting all this image data in addition to the WIT dataset itself and will provide colab notebooks. Further, the competitors will have access to a discussion forum in Kaggle in order to share code and collaborate. This enables anyone interested in multimodality to get started and run experiments easily. We are excited and looking forward to what will result from the WIT dataset and the Wikipedia images in the Kaggle platform.
Conclusion We believe that the WIT dataset will aid researchers in building better multimodal multilingual models and in identifying better learning and representation techniques, ultimately leading to improved Machine Learning models in real-world tasks over visio-linguistic data. For any questions, please contact wit-dataset@google.com. We would love to hear about how you are using the WIT dataset.
Acknowledgements We would like to thank our co-authors in Google Research: Jiecao Chen, Michael Bendersky and Marc Najork. We thank Beer Changpinyo, Corinna Cortes, Joshua Gang, Chao Jia, Ashwin Kakarla, Mike Lee, Zhen Li, Piyush Sharma, Radu Soricut, Ashish Vaswani, Yinfei Yang, and our reviewers for their insightful feedback and comments.
We thank Miriam Redi and Leila Zia from Wikimedia Research for collaborating with us on the competition and providing image pixels and image embedding data. We thank Addison Howard and Walter Reade for helping us host this competition in Kaggle. We also thank Diane Larlus (Naver Labs Europe (NLE)), Yannis Kalantidis (NLE), Stéphane Clinchant (NLE), Tiziano Piccardi Ph.D. student at EPFL, Lucie-Aimée Kaffee PhD student at University of Southampton and Yacine Jernite (Hugging Face) for their valuable contribution towards the competition.
As neural network models and training data size grow, training efficiency is becoming an important focus for deep learning. For example, GPT-3 demonstrates remarkable capability in few-shot learning, but it requires weeks of training with thousands of GPUs, making it difficult to retrain or improve. What if, instead, one could design neural networks that were smaller and faster, yet still more accurate?
In this post, we introduce two families of models for image recognition that leverage neural architecture search, and a principled design methodology based on model capacity and generalization. The first is EfficientNetV2 (accepted at ICML 2021), which consists of convolutional neural networks that aim for fast training speed for relatively small-scale datasets, such as ImageNet1k (with 1.28 million images). The second family is CoAtNet, which are hybrid models that combine convolution and self-attention, with the goal of achieving higher accuracy on large-scale datasets, such as ImageNet21 (with 13 million images) and JFT (with billions of images). Compared to previous results, our models are 4-10x faster while achieving new state-of-the-art 90.88% top-1 accuracy on the well-established ImageNet dataset. We are also releasing the source code and pretrained models on the Google AutoML github.
EfficientNetV2: Smaller Models and Faster Training EfficientNetV2 is based upon the previous EfficientNet architecture. To improve upon the original, we systematically studied the training speed bottlenecks on modern TPUs/GPUs and found: (1) training with very large image sizes results in higher memory usage and thus is often slower on TPUs/GPUs; (2) the widely used depthwise convolutions are inefficient on TPUs/GPUs, because they exhibit low hardware utilization; and (3) the commonly used uniform compound scaling approach, which scales up every stage of convolutional networks equally, is sub-optimal. To address these issues, we propose both a training-aware neural architecture search (NAS), in which the training speed is included in the optimization goal, and a scaling method that scales different stages in a non-uniform manner.
The training-aware NAS is based on the previous platform-aware NAS, but unlike the original approach, which mostly focuses on inference speed, here we jointly optimize model accuracy, model size, and training speed. We also extend the original search space to include more accelerator-friendly operations, such as FusedMBConv, and simplify the search space by removing unnecessary operations, such as average pooling and max pooling, which are never selected by NAS. The resulting EfficientNetV2 networks achieve improved accuracy over all previous models, while being much faster and up to 6.8x smaller.
To further speed up the training process, we also propose an enhanced method of progressive learning, which gradually changes image size and regularization magnitude during training. Progressive training has been used in image classification, GANs, and language models. This approach focuses on image classification, but unlike previous approaches that often trade accuracy for improved training speed, can slightly improve the accuracy while also significantly reducing training time. The key idea in our improved approach is to adaptively change regularization strength, such as dropout ratio or data augmentation magnitude, according to the image size. For the same network, small image size leads to lower network capacity and thus requires weak regularization; vice versa, a large image size requires stronger regularization to combat overfitting.
We evaluate the EfficientNetV2 models on ImageNet and a few transfer learning datasets, such as CIFAR-10/100, Flowers, and Cars. On ImageNet, EfficientNetV2 significantly outperforms previous models with about 5–11x faster training speed and up to 6.8x smaller model size, without any drop in accuracy.
CoAtNet: Fast and Accurate Models for Large-Scale Image Recognition While EfficientNetV2 is still a typical convolutional neural network, recent studies on Vision Transformer (ViT) have shown that attention-based transformer models could perform better than convolutional neural networks on large-scale datasets like JFT-300M. Inspired by this observation, we further expand our study beyond convolutional neural networks with the aim of finding faster and more accurate vision models.
In “CoAtNet: Marrying Convolution and Attention for All Data Sizes”, we systematically study how to combine convolution and self-attention to develop fast and accurate neural networks for large-scale image recognition. Our work is based on an observation that convolution often has better generalization (i.e., the performance gap between training and evaluation) due to its inductive bias, while self-attention tends to have greater capacity (i.e., the ability to fit large-scale training data) thanks to its global receptive field. By combining convolution and self-attention, our hybrid models can achieve both better generalization and greater capacity.
We observe two key insights from our study: (1) depthwise convolution and self-attention can be naturally unified via simple relative attention, and (2) vertically stacking convolution layers and attention layers in a way that considers their capacity and computation required in each stage (resolution) is surprisingly effective in improving generalization, capacity and efficiency. Based on these insights, we have developed a family of hybrid models with both convolution and attention, named CoAtNets (pronounced “coat” nets). The following figure shows the overall CoAtNet network architecture:
CoAtNet models consistently outperform ViT models and its variants across a number of datasets, such as ImageNet1K, ImageNet21K, and JFT. When compared to convolutional networks, CoAtNet exhibits comparable performance on a small-scale dataset (ImageNet1K) and achieves substantial gains as the data size increases (e.g. on ImageNet21K and JFT).
We also evaluated CoAtNets on the large-scale JFT dataset. To reach a similar accuracy target, CoAtNet trains about 4x faster than previous ViT models and more importantly, achieves a new state-of-the-art top-1 accuracy on ImageNet of 90.88%.
Conclusion and Future Work In this post, we introduce two families of neural networks, named EfficientNetV2 and CoAtNet, which achieve state-of-the-art performance on image recognition. All EfficientNetV2 models are open sourced and the pretrained models are also available on the TFhub. CoAtNet models will also be open-sourced soon. We hope these new neural networks can benefit the research community and the industry. In the future we plan to further optimize these models and apply them to new tasks, such as zero-shot learning and self-supervised learning, which often require fast models with high capacity.
Acknowledgements Special thanks to our co-authors Hanxiao Liu and Quoc Le. We also thank the Google Research, Brain Team and the open source contributors.
Instance segmentation is the task of grouping pixels in an image into instances of individual things, and identifying those things with a class label (countable objects such as people, animals, cars, etc., and assigning unique identifiers to each, e.g., car_1 and car_2). As a core computer vision task, it is critical to many downstream applications, such as self-driving cars, robotics, medical imaging, and photo editing. In recent years, deep learning has made significant strides in solving the instance segmentation problem with architectures like Mask R-CNN. However, these methods rely on collecting a large labeled instance segmentation dataset. But unlike bounding box labels, which can be collected in 7 seconds per instance with methods like Extreme clicking, collecting instance segmentation labels (called “masks”) can take up to 80 seconds per instance, an effort that is costly and creates a high barrier to entry for this research. And a related task, pantopic segmentation, requires even more labeled data.
The partially supervised instance segmentation setting, where only a small set of classes are labeled with instance segmentation masks and the remaining (majority of) classes are labeled only with bounding boxes, is an approach that has the potential to reduce the dependence on manually-created mask labels, thereby significantly lowering the barriers to developing an instance segmentation model. However this partially supervised approach also requires a stronger form of model generalization to handle novel classes not seen at training time—e.g., training with only animal masks and then tasking the model to produce accurate instance segmentations for buildings or plants. Further, naïve approaches, such as training a class-agnostic Mask R-CNN, while ignoring mask losses for any instances that don’t have mask labels, have not worked well. For example, on the typical “VOC/Non-VOC” benchmark, where one trains on masks for a subset of 20 classes in COCO (called “seen classes”) and is tested on the remaining 60 classes (called “unseen classes”), a typical Mask R-CNN with Resnet-50 backbone gets to only ~18% mask mAP (mean Average Precision, higher is better) on unseen classes, whereas when fully supervised it can achieve a much higher >34% mask mAP on the same set.
In “The surprising impact of mask-head architecture on novel class segmentation”, to be presented at ICCV 2021, we identify the main culprits for Mask R-CNN’s poor performance on novel classes and propose two easy-to-implement fixes (one training protocol fix, one mask-head architecture fix) that work in tandem to close the gap to fully supervised performance. We show that our approach applies generally to crop-then-segment models, i.e., a Mask R-CNN or Mask R-CNN-like architecture that computes a feature representation of the entire image and then subsequently passes per-instance crops to a second-stage mask prediction network—also called a mask-head network. Putting our findings together, we propose a Mask R-CNN–based model that improves over the current state-of-the-art by a significant 4.7% mask mAP without requiring more complex auxiliary loss functions, offline trained priors, or weight transfer functions proposed by previous work. We have also open sourced the code bases for two versions of the model, called Deep-MAC and Deep-MARC, and published a colab to interactively produce masks like the video demo below.
Impact of Cropping Methodology in Partially Supervised Settings An important step of crop-then-segment models is cropping—Mask R-CNN is trained by cropping a feature map as well as the ground truth mask to a bounding box corresponding to each instance. These cropped features are passed to another neural network (called a mask-head network) that computes a final mask prediction, which is then compared against the ground truth crop in the mask loss function. There are two choices for cropping: (1) cropping directly to the ground truth bounding box of an instance, or (2) cropping to bounding boxes predicted by the model (called, proposals). At test time, cropping is always performed with proposals as ground truth boxes are not assumed to be available.
Typical Mask R-CNN implementations pass both types of crops to the mask head. However, this choice has traditionally been considered an unimportant implementation detail, because it does not affect performance significantly in the fully supervised setting. In contrast, for partially supervised settings, we find that cropping methodology plays a significant role—while cropping exclusively to ground truth boxes during training doesn’t change the results significantly in the fully supervised setting, it has a surprising and dramatic positive impact in the partially supervised setting, performing significantly better on unseen classes.
Unlocking the Full Generalization Potential of the Mask Head Even more surprisingly, the above approach unlocks a novel phenomenon—with cropping-to-ground truth enabled during training, the mask head of Mask R-CNN takes on a disproportionate role in the ability of the model to generalize to unseen classes. As an example, in the following figure, we compare models that all have cropping-to-ground-truth enabled, but different out-of-the-box mask-head architectures on a parking meter, cell phone, and pizza (classes unseen during training).
Particularly notable is that these differences between mask-head architectures are not as obvious in the fully supervised setting. Incidentally, this may explain why previous works in instance segmentation have almost exclusively used shallow (i.e., low number of layers) mask heads, as there has been no benefit to the added complexity. Below we compare the mask mAP of three different mask-head architectures on seen versus unseen classes. All three models do equally well on the set of seen classes, but the deep hourglass mask heads stand out when applied to unseen classes. We find hourglass mask heads to be the best among the architectures we tried and we use hourglass mask heads with 50 or more layers to get the best results.
Finally, we show that our findings are general, holding for a variety of backbones (e.g., ResNet, SpineNet, Hourglass) and detector architectures including anchor-based and anchor-free detectors and even when there is no detector at all.
Putting It Together To achieve the best result, we combined the above findings: We trained a Mask R-CNN model with cropping-to-ground-truth enabled and a deep Hourglass-52 mask head with a SpineNet backbone on high resolution images (1280x1280). We call this model Deep-MARC (Deep Mask heads Above R-CNN). Without using any offline training or other hand-crafted priors, Deep-MARC exceeds previous state-of-the-art models by > 4.5% (absolute) mask mAP. Demonstrating the general nature of this approach, we also see strong results with a CenterNet-based (as opposed to Mask R-CNN-based) model (called Deep-MAC), which also exceeds the previous state of the art.
Conclusion We develop instance segmentation models that are able to generalize to classes that were not part of the training set. We highlight the role of two key ingredients that can be applied to any crop-then-segment model (such as Mask R-CNN): (1) cropping-to-ground truth boxes during training, and (2) strong mask-head architectures. While neither of these ingredients have a large impact on the classes for which masks are available during training, employing both leads to significant improvement on novel classes for which masks are not available during training. Moreover, these ingredients are sufficient for achieving state-of-the-art-performance on the partially-supervised COCO benchmark. Finally, our findings are general and may also have implications for related tasks, such as panoptic segmentation and pose estimation.
Acknowledgements We thank our co-authors Zhichao Lu, Siyang Li, and Vivek Rathod. We thank David Ross and our anonymous ICCV reviewers for their comments which played a big part in improving this research.
Dancing is a universal language found in nearly all cultures, and is an outlet many people use to express themselves on contemporary media platforms today. The ability to dance by composing movement patterns that align to music beats is a fundamental aspect of human behavior. However, dancing is a form of art that requires practice. In fact, professional training is often required to equip a dancer with a rich repertoire of dance motions needed to create expressive choreography. While this process is difficult for people, it is even more challenging for a machine learning (ML) model, because the task requires the ability to generate a continuous motion with high kinematic complexity, while capturing the non-linear relationship between the movements and the accompanying music.
In “AI Choreographer: Music-Conditioned 3D Dance Generation with AIST++”, presented at ICCV 2021, we propose a full-attention cross-modal Transformer (FACT) model can mimic and understand dance motions, and can even enhance a person’s ability to choreograph dance. Together with the model, we released a large-scale, multi-modal 3D dance motion dataset, AIST++, which contains 5.2 hours of 3D dance motion in 1408 sequences, covering 10 dance genres, each including multi-view videos with known camera poses. Through extensive user studies on AIST++, we find that the FACT model outperforms recent state-of-the-art methods, both qualitatively and quantitatively.
We generate the proposed 3D motion dataset from the existing AIST Dance Video Database — a collection of videos of dance with musical accompaniment, but without any 3D information. AIST contains 10 dance genres: Old School (Break, Pop, Lock and Waack) and New School (Middle Hip-Hop, LA-style Hip-Hop, House, Krump, Street Jazz and Ballet Jazz). Although it contains multi-view videos of dancers, these cameras are not calibrated.
For our purposes, we recovered the camera calibration parameters and the 3D human motion in terms of parameters used by the widely used SMPL 3D model. The resulting database, AIST++, is a large-scale, 3D human dance motion dataset that contains a wide variety of 3D motion, paired with music. Each frame includes extensive annotations:
The motions are equally distributed among all 10 dance genres, covering a wide variety of music tempos in beat per minute (BPM). Each genre of dance contains 85% basic movements and 15% advanced movements (longer choreographies freely designed by the dancers).
The AIST++ dataset also contains multi-view synchronized image data, making it useful for other research directions, such as 2D/3D pose estimation. To our knowledge, AIST++ is the largest 3D human dance dataset with 1408 sequences, 30 subjects and 10 dance genres, and with both basic and advanced choreographies.
Because AIST is an instructional database, it records multiple dancers following the same choreography for different music with varying BPM, a common practice in dance. This posits a unique challenge in cross-modal sequence-to-sequence generation as the model needs to learn the one-to-many mapping between audio and motion. We carefully construct non-overlapping train and test subsets on AIST++ to ensure neither choreography nor music is shared across the subsets.
Full Attention Cross-Modal Transformer (FACT) Model Using this data, we train the FACT model to generate 3D dance from music. The model begins by encoding seed motion and audio inputs using separate motion and audio transformers. The embeddings are then concatenated and sent to a cross-modal transformer, which learns the correspondence between both modalities and generates N future motion sequences. These sequences are then used to train the model in a self-supervised manner. All three transformers are jointly learned end-to-end. At test time, we apply this model in an autoregressive framework, where the predicted motion serves as the input to the next generation step. As a result, the FACT model is capable of generating long range dance motion frame-by-frame.
FACT involves three key design choices that are critical for producing realistic 3D dance motion from music.
Results We evaluate the performance based on three metrics:
Motion Quality: We calculate the Frechet Inception Distance (FID) between the real dance motion sequences in the AIST++ test set and 40 model generated motion sequences, each with 1200 frames (20 secs). We denote the FID based on the geometric and kinetic features as FIDg and FIDk, respectively.
Generation Diversity: Similar to prior work, to evaluate the model’s ability to generate divers dance motions, we calculate the average Euclidean distance in the feature space across 40 generated motions on the AIST++ test set, again comparing geometric feature space (Distg) and in the kinetic feature space (Distk).
Motion-Music Correlation: Because there is no well-designed metric to measure the correlation between input music (music beats) and generated 3D motion (kinematic beats), we propose a novel metric, called Beat Alignment Score (BeatAlign).
Quantitative Evaluation We compare the performance of FACT on each of these metrics to that of other state-of-the-art methods.
We also perceptually evaluate the motion-music correlation with a user study in which each participant is asked to watch 10 videos showing one of our results and one random counterpart, and then select which dancer is more in sync with the music. The study consisted of 30 participants, ranging from professional dancers to people who rarely dance. Compared to each baseline, 81% prefered the FACT model output to that of Li et al., 71% prefered FACT to Dancenet, and 77% prefered it Dance Revolution. Interestingly, 75% of participants preferred the unpaired AIST++ dance motion to that generated by FACT, which is unsurprising since the original dance captures are highly expressive.
Qualitative Results Compared with prior methods like DanceNet (left) and Li et. al. (middle), 3D dance generated using the FACT model (right) is more realistic and better correlated with input music.
More generated 3D dances using the FACT model.
Conclusion and Discussion We present a model that can not only learn the audio-motion correspondence, but also can generate high quality 3D motion sequences conditioned on music. Because generating 3D movement from music is a nascent area of study, we hope our work will pave the way for future cross-modal audio to 3D motion generation. We are also releasing AIST++, the largest 3D human dance dataset to date. This proposed, multi-view, multi-genre, cross-modal 3D motion dataset can not only help research in the conditional 3D motion generation research but also human understanding research in general. We are releasing the code in our GitHub repository and the trained model here.
While our results show a promising direction in this problem of music conditioned 3D motion generation, there are more to be explored. First, our approach is kinematic-based and we do not reason about physical interactions between the dancer and the floor. Therefore the global translation can lead to artifacts, such as foot sliding and floating. Second, our model is currently deterministic. Exploring how to generate multiple realistic dances per music is an exciting direction.
Acknowledgements We gratefully acknowledge the contribution of other co-authors, including Ruilong Li and David Ross. We thank Chen Sun, Austin Myers, Bryan Seybold and Abhijit Kundu for helpful discussions. We thank Emre Aksan and Jiaman Li for sharing their code. We also thank Kevin Murphy for the early attempts in this direction, as well as Peggy Chi and Pan Chen for the help on user study experiments. This work would not have been possible without the AIST Dance Video Database (https://aistdancedb.ongaaccel.jp), so we would like to especially thank its creators for their valuable contribution to the community.
