视觉-语言-动作（VLA）

Large-scale transformer models bring significant improvements for various downstream vision language tasks with a unified architecture. The performance improvements come with increasing model size, resulting in slow inference speed and increased cost for severing. While some certain predictions benefit from the full complexity of the large-scale model, not all of input need the same amount of computation to conduct, potentially leading to computation resource waste. To handle this challenge, early exiting is proposed to adaptively allocate computational power in term of input complexity to improve inference efficiency. The existing early exiting strategies usually adopt output confidence based on intermediate layers as a proxy of input complexity to incur the decision of skipping following layers. However, such strategies cannot apply to encoder in the widely-used unified architecture with both encoder and decoder due to difficulty of output confidence estimation in the encoder. It is suboptimal in term of saving computation power to ignore the early exiting in encoder component. To handle this challenge, we propose a novel early exiting strategy for unified visual language models, which allows dynamically skip the layers in encoder and decoder simultaneously in term of input layer-wise similarities with multiple times of early exiting, namely MuE. By decomposing the image and text modalities in the encoder, MuE is flexible and can skip different layers in term of modalities, advancing the inference efficiency while minimizing performance drop. Experiments on the SNLI-VE and MS COCO datasets show that the proposed approach MuE can reduce inference time by up to 50% and 40% while maintaining 99% and 96% performance respectively.

Recently vision transformer models have become prominent models for a range of tasks. These models, however, usually suffer from intensive computational costs and heavy memory requirements, making them impractical for deployment on edge platforms. Recent studies have proposed to prune transformers in an unexplainable manner, which overlook the relationship between internal units of the model and the target class, thereby leading to inferior performance. To alleviate this problem, we propose a novel explainable pruning framework dubbed X-Pruner, which is designed by considering the explainability of the pruning criterion. Specifically, to measure each prunable unit's contribution to predicting each target class, a novel explainability-aware mask is proposed and learned in an end-to-end manner. Then, to preserve the most informative units and learn the layer-wise pruning rate, we adaptively search the layer-wise threshold that differentiates between unpruned and pruned units based on their explainability-aware mask values. To verify and evaluate our method, we apply the X-Pruner on representative transformer models including the DeiT and Swin Transformer. Comprehensive simulation results demonstrate that the proposed X-Pruner outperforms the state-of-the-art black-box methods with significantly reduced computational costs and slight performance degradation.

3D reconstruction is a fundamental problem in computer vision, and the task is especially challenging when the object to reconstruct is partially or fully occluded. We introduce a method that uses the shadows cast by an unobserved object in order to infer the possible 3D volumes under occlusion. We create a differentiable image formation model that allows us to jointly infer the 3D shape of an object, its pose, and the position of a light source. Since the approach is end-to-end differentiable, we are able to integrate learned priors of object geometry in order to generate realistic 3D shapes of different object categories. Experiments and visualizations show that the method is able to generate multiple possible solutions that are consistent with the observation of the shadow. Our approach works even when the position of the light source and object pose are both unknown. Our approach is also robust to real-world images where ground-truth shadow mask is unknown.

Formula-driven supervised learning (FDSL) has been shown to be an effective method for pre-training vision transformers, where ExFractalDB-21k was shown to exceed the pre-training effect of ImageNet-21k. These studies also indicate that contours mattered more than textures when pre-training vision transformers. However, the lack of a systematic investigation as to why these contour-oriented synthetic datasets can achieve the same accuracy as real datasets leaves much room for skepticism. In the present work, we develop a novel methodology based on circular harmonics for systematically investigating the design space of contour-oriented synthetic datasets. This allows us to efficiently search the optimal range of FDSL parameters and maximize the variety of synthetic images in the dataset, which we found to be a critical factor. When the resulting new dataset VisualAtom-21k is used for pre-training ViT-Base, the top-1 accuracy reached 83.7% when fine-tuning on ImageNet-1k. This is only 0.5% difference from the top-1 accuracy (84.2%) achieved by the JFT-300M pre-training, even though the scale of images is 1/14. Unlike JFT-300M which is a static dataset, the quality of synthetic datasets will continue to improve, and the current work is a testament to this possibility. FDSL is also free of the common issues associated with real images, e.g. privacy/copyright issues, labeling costs/errors, and ethical biases.

We propose Mask Auto-Labeler (MAL), a high-quality Transformer-based mask auto-labeling framework for instance segmentation using only box annotations. MAL takes box-cropped images as inputs and conditionally generates their mask pseudo-labels.We show that Vision Transformers are good mask auto-labelers. Our method significantly reduces the gap between auto-labeling and human annotation regarding mask quality. Instance segmentation models trained using the MAL-generated masks can nearly match the performance of their fully-supervised counterparts, retaining up to 97.4% performance of fully supervised models. The best model achieves 44.1% mAP on COCO instance segmentation (test-dev 2017), outperforming state-of-the-art box-supervised methods by significant margins. Qualitative results indicate that masks produced by MAL are, in some cases, even better than human annotations.

This paper simultaneously addresses three limitations associated with conventional skeleton-based action recognition; skeleton detection and tracking errors, poor variety of the targeted actions, as well as person-wise and frame-wise action recognition. A point cloud deep-learning paradigm is introduced to the action recognition, and a unified framework along with a novel deep neural network architecture called Structured Keypoint Pooling is proposed. The proposed method sparsely aggregates keypoint features in a cascaded manner based on prior knowledge of the data structure (which is inherent in skeletons), such as the instances and frames to which each keypoint belongs, and achieves robustness against input errors. Its less constrained and tracking-free architecture enables time-series keypoints consisting of human skeletons and nonhuman object contours to be efficiently treated as an input 3D point cloud and extends the variety of the targeted action. Furthermore, we propose a Pooling-Switching Trick inspired by Structured Keypoint Pooling. This trick switches the pooling kernels between the training and inference phases to detect person-wise and frame-wise actions in a weakly supervised manner using only video-level action labels. This trick enables our training scheme to naturally introduce novel data augmentation, which mixes multiple point clouds extracted from different videos. In the experiments, we comprehensively verify the effectiveness of the proposed method against the limitations, and the method outperforms state-of-the-art skeleton-based action recognition and spatio-temporal action localization methods.

Despite the remarkable success of foundation models, their task-specific fine-tuning paradigm makes them inconsistent with the goal of general perception modeling. The key to eliminating this inconsistency is to use generalist models for general task modeling. However, existing attempts at generalist models are inadequate in both versatility and performance. In this paper, we propose Uni-Perceiver v2, which is the first generalist model capable of handling major large-scale vision and vision-language tasks with competitive performance. Specifically, images are encoded as general region proposals, while texts are encoded via a Transformer-based language model. The encoded representations are transformed by a task-agnostic decoder. Different tasks are formulated as a unified maximum likelihood estimation problem. We further propose an effective optimization technique named Task-Balanced Gradient Normalization to ensure stable multi-task learning with an unmixed sampling strategy, which is helpful for tasks requiring large batch-size training. After being jointly trained on various tasks, Uni-Perceiver v2 is capable of directly handling downstream tasks without any task-specific adaptation. Results show that Uni-Perceiver v2 outperforms all existing generalist models in both versatility and performance. Meanwhile, compared with the commonly-recognized strong baselines that require tasks-specific fine-tuning, Uni-Perceiver v2 achieves competitive performance on a broad range of vision and vision-language tasks.

This paper proposes a novel method for vision-based metric cross-view geolocalization (CVGL) that matches the camera images captured from a ground-based vehicle with an aerial image to determine the vehicle's geo-pose. Since aerial images are globally available at low cost, they represent a potential compromise between two established paradigms of autonomous driving, i.e. using expensive high-definition prior maps or relying entirely on the sensor data captured at runtime. We present an end-to-end differentiable model that uses the ground and aerial images to predict a probability distribution over possible vehicle poses. We combine multiple vehicle datasets with aerial images from orthophoto providers on which we demonstrate the feasibility of our method. Since the ground truth poses are often inaccurate w.r.t. the aerial images, we implement a pseudo-label approach to produce more accurate ground truth poses and make them publicly available. While previous works require training data from the target region to achieve reasonable localization accuracy (i.e. same-area evaluation), our approach overcomes this limitation and outperforms previous results even in the strictly more challenging cross-area case. We improve the previous state-of-the-art by a large margin even without ground or aerial data from the test region, which highlights the model's potential for global-scale application. We further integrate the uncertainty-aware predictions in a tracking framework to determine the vehicle's trajectory over time resulting in a mean position error on KITTI-360 of 0.78m.

Weakly-supervised temporal action localization aims to detect action boundaries in untrimmed videos with only video-level annotations. Most existing schemes detect temporal regions that are most responsive to video-level classification, but they overlook the semantic consistency between frames. In this paper, we hypothesize that snippets with similar representations should be considered as the same action class despite the absence of supervision signals on each snippet. To this end, we devise a learnable dictionary where entries are the class centroids of the corresponding action categories. The representations of snippets identified as the same action category are induced to be close to the same class centroid, which guides the network to perceive the semantics of frames and avoid unreasonable localization. Besides, we propose a two-stream framework that integrates the attention mechanism and the multiple-instance learning strategy to extract fine-grained clues and salient features respectively. Their complementarity enables the model to refine temporal boundaries. Finally, the developed model is validated on the publicly available THUMOS-14 and ActivityNet-1.3 datasets, where substantial experiments and analyses demonstrate that our model achieves remarkable advances over existing methods.

Vision Transformers (ViTs) have demonstrated the state-of-the-art performance in various vision-related tasks. The success of ViTs motivates adversaries to perform backdoor attacks on ViTs. Although the vulnerability of traditional CNNs to backdoor attacks is well-known, backdoor attacks on ViTs are seldom-studied. Compared to CNNs capturing pixel-wise local features by convolutions, ViTs extract global context information through patches and attentions. Naively transplanting CNN-specific backdoor attacks to ViTs yields only a low clean data accuracy and a low attack success rate. In this paper, we propose a stealth and practical ViT-specific backdoor attack TrojViT. Rather than an area-wise trigger used by CNN-specific backdoor attacks, TrojViT generates a patch-wise trigger designed to build a Trojan composed of some vulnerable bits on the parameters of a ViT stored in DRAM memory through patch salience ranking and attention-target loss. TrojViT further uses parameter distillation to reduce the bit number of the Trojan. Once the attacker inserts the Trojan into the ViT model by flipping the vulnerable bits, the ViT model still produces normal inference accuracy with benign inputs. But when the attacker embeds a trigger into an input, the ViT model is forced to classify the input to a predefined target class. We show that flipping only few vulnerable bits identified by TrojViT on a ViT model using the well-known RowHammer can transform the model into a backdoored one. We perform extensive experiments of multiple datasets on various ViT models. TrojViT can classify 99.64% of test images to a target class by flipping 345 bits on a ViT for ImageNet.

In this paper, we present a one-stage framework TriDet for temporal action detection. Existing methods often suffer from imprecise boundary predictions due to the ambiguous action boundaries in videos. To alleviate this problem, we propose a novel Trident-head to model the action boundary via an estimated relative probability distribution around the boundary. In the feature pyramid of TriDet, we propose a Scalable-Granularity Perception (SGP) layer to aggregate information across different temporal granularities, which is much more efficient than the recent transformer-based feature pyramid. Benefiting from the Trident-head and the SGP-based feature pyramid, TriDet achieves state-of-the-art performance on three challenging benchmarks: THUMOS14, HACS and EPIC-KITCHEN 100, with lower computational costs, compared to previous methods. For example, TriDet hits an average mAP of 69.3% on THUMOS14, outperforming the previous best by 2.5%, but with only 74.6% of its latency.

Vision transformers (ViTs) have been successfully deployed in a variety of computer vision tasks, but they are still vulnerable to adversarial samples. Transfer-based attacks use a local model to generate adversarial samples and directly transfer them to attack a target black-box model. The high efficiency of transfer-based attacks makes it a severe security threat to ViT-based applications. Therefore, it is vital to design effective transfer-based attacks to identify the deficiencies of ViTs beforehand in security-sensitive scenarios. Existing efforts generally focus on regularizing the input gradients to stabilize the updated direction of adversarial samples. However, the variance of the back-propagated gradients in intermediate blocks of ViTs may still be large, which may make the generated adversarial samples focus on some model-specific features and get stuck in poor local optima. To overcome the shortcomings of existing approaches, we propose the Token Gradient Regularization (TGR) method. According to the structural characteristics of ViTs, TGR reduces the variance of the back-propagated gradient in each internal block of ViTs in a token-wise manner and utilizes the regularized gradient to generate adversarial samples. Extensive experiments on attacking both ViTs and CNNs confirm the superiority of our approach. Notably, compared to the state-of-the-art transfer-based attacks, our TGR offers a performance improvement of 8.8 % on average.

Although deep neural networks (DNNs) have shown great successes in computer vision tasks, they are vulnerable to perturbations on inputs, and there exists a trade-off between the natural accuracy and robustness to such perturbations, which is mainly caused by the existence of robust non-predictive features and non-robust predictive features. Recent empirical analyses find Vision Transformers (ViTs) are inherently robust to various kinds of perturbations, but the aforementioned trade-off still exists for them. In this work, we propose Trade-off between Robustness and Accuracy of Vision Transformers (TORA-ViTs), which aims to efficiently transfer ViT models pretrained on natural tasks for both accuracy and robustness. TORA-ViTs consist of two major components, including a pair of accuracy and robustness adapters to extract predictive and robust features, respectively, and a gated fusion module to adjust the trade-off. The gated fusion module takes outputs of a pretrained ViT block as queries and outputs of our adapters as keys and values, and tokens from different adapters at different spatial locations are compared with each other to generate attention scores for a balanced mixing of predictive and robust features. Experiments on ImageNet with various robust benchmarks show that our TORA-ViTs can efficiently improve the robustness of naturally pretrained ViTs while maintaining competitive natural accuracy. Our most balanced setting (TORA-ViTs with lambda = 0.5) can maintain 83.7% accuracy on clean ImageNet and reach 54.7% and 38.0% accuracy under FGSM and PGD white-box attacks, respectively. In terms of various ImageNet variants, it can reach 39.2% and 56.3% accuracy on ImageNet-A and ImageNet-R and reach 34.4% mCE on ImageNet-C.

We propose Token Turing Machines (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which summarise the previous history (i.e., frames). This memory is efficiently addressed, read and written using a Transformer as the processing unit/controller at each step. The model's memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efficiently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential visual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token_turing.

Semi-Supervised Learning can be more beneficial for the video domain compared to images because of its higher annotation cost and dimensionality. Besides, any video understanding task requires reasoning over both spatial and temporal dimensions. In order to learn both the static and motion related features for the semi-supervised action recognition task, existing methods rely on hard input inductive biases like using two-modalities (RGB and Optical-flow) or two-stream of different playback rates. Instead of utilizing unlabeled videos through diverse input streams, we rely on self-supervised video representations, particularly, we utilize temporally-invariant and temporally-distinctive representations. We observe that these representations complement each other depending on the nature of the action. Based on this observation, we propose a student-teacher semi-supervised learning framework, TimeBalance, where we distill the knowledge from a temporally-invariant and a temporally-distinctive teacher. Depending on the nature of the unlabeled video, we dynamically combine the knowledge of these two teachers based on a novel temporal similarity-based reweighting scheme. Our method achieves state-of-the-art performance on three action recognition benchmarks: UCF101, HMDB51, and Kinetics400. Code: https://github.com/DAVEISHAN/TimeBalance.

End-to-end autonomous driving has made impressive progress in recent years. Existing methods usually adopt the decoupled encoder-decoder paradigm, where the encoder extracts hidden features from raw sensor data, and the decoder outputs the ego-vehicle's future trajectories or actions. Under such a paradigm, the encoder does not have access to the intended behavior of the ego agent, leaving the burden of finding out safety-critical regions from the massive receptive field and inferring about future situations to the decoder. Even worse, the decoder is usually composed of several simple multi-layer perceptrons (MLP) or GRUs while the encoder is delicately designed (e.g., a combination of heavy ResNets or Transformer). Such an imbalanced resource-task division hampers the learning process. In this work, we aim to alleviate the aforementioned problem by two principles: (1) fully utilizing the capacity of the encoder; (2) increasing the capacity of the decoder. Concretely, we first predict a coarse-grained future position and action based on the encoder features. Then, conditioned on the position and action, the future scene is imagined to check the ramification if we drive accordingly. We also retrieve the encoder features around the predicted coordinate to obtain fine-grained information about the safety-critical region. Finally, based on the predicted future and the retrieved salient feature, we refine the coarse-grained position and action by predicting its offset from ground-truth. The above refinement module could be stacked in a cascaded fashion, which extends the capacity of the decoder with spatial-temporal prior knowledge about the conditioned future. We conduct experiments on the CARLA simulator and achieve state-of-the-art performance in closed-loop benchmarks. Extensive ablation studies demonstrate the effectiveness of each proposed module. Code and models are available at https://github.com/opendrivelab/ThinkTwice.

In this paper we introduce a rule-based, compositional, and hierarchical modeling of action using Therbligs as our atoms. Introducing these atoms provides us with a consistent, expressive, contact-centered representation of action. Over the atoms we introduce a differentiable method of rule-based reasoning to regularize for logical consistency. Our approach is complementary to other approaches in that the Therblig-based representations produced by our architecture augment rather than replace existing architectures' representations. We release the first Therblig-centered annotations over two popular video datasets - EPIC Kitchens 100 and 50-Salads. We also broadly demonstrate benefits to adopting Therblig representations through evaluation on the following tasks: action segmentation, action anticipation, and action recognition - observing an average 10.5%/7.53%/6.5% relative improvement, respectively, over EPIC Kitchens and an average 8.9%/6.63%/4.8% relative improvement, respectively, over 50 Salads. Code and data will be made publicly available.

Early action prediction deals with inferring the ongoing action from partially-observed videos, typically at the outset of the video. We propose a bottleneck-based attention model that captures the evolution of the action, through progressive sampling over fine-to-coarse scales. Our proposed Temporal Progressive (TemPr) model is composed of multiple attention towers, one for each scale. The predicted action label is based on the collective agreement considering confidences of these towers. Extensive experiments over four video datasets showcase state-of-the-art performance on the task of Early Action Prediction across a range of encoder architectures. We demonstrate the effectiveness and consistency of TemPr through detailed ablations.

Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision & Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to 15% in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model. Our code and pretrained models are available at: https://github.com/SivanDoveh/TSVLC