CVPR 2023 Papers

Optical flow is an indispensable building block for various important computer vision tasks, including motion estimation, object tracking, and disparity measurement. In this work, we propose TransFlow, a pure transformer architecture for optical flow estimation. Compared to dominant CNN-based methods, TransFlow demonstrates three advantages. First, it provides more accurate correlation and trustworthy matching in flow estimation by utilizing spatial self-attention and cross-attention mechanisms between adjacent frames to effectively capture global dependencies; Second, it recovers more compromised information (e.g., occlusion and motion blur) in flow estimation through long-range temporal association in dynamic scenes; Third, it enables a concise self-learning paradigm and effectively eliminate the complex and laborious multi-stage pre-training procedures. We achieve the state-of-the-art results on the Sintel, KITTI-15, as well as several downstream tasks, including video object detection, interpolation and stabilization. For its efficacy, we hope TransFlow could serve as a flexible baseline for optical flow estimation.

Multi-person pose forecasting remains a challenging problem, especially in modeling fine-grained human body interaction in complex crowd scenarios. Existing methods typically represent the whole pose sequence as a temporal series, yet overlook interactive influences among people based on skeletal body parts. In this paper, we propose a novel Trajectory-Aware Body Interaction Transformer (TBIFormer) for multi-person pose forecasting via effectively modeling body part interactions. Specifically, we construct a Temporal Body Partition Module that transforms all the pose sequences into a Multi-Person Body-Part sequence to retain spatial and temporal information based on body semantics. Then, we devise a Social Body Interaction Self-Attention (SBI-MSA) module, utilizing the transformed sequence to learn body part dynamics for inter- and intra-individual interactions. Furthermore, different from prior Euclidean distance-based spatial encodings, we present a novel and efficient Trajectory-Aware Relative Position Encoding for SBI-MSA to offer discriminative spatial information and additional interactive clues. On both short- and long-term horizons, we empirically evaluate our framework on CMU-Mocap, MuPoTS-3D as well as synthesized datasets (6 10 persons), and demonstrate that our method greatly outperforms the state-of-the-art methods.

Neural networks are often biased to spuriously correlated features that provide misleading statistical evidence that does not generalize. This raises an interesting question: "Does an optimal unbiased functional subnetwork exist in a severely biased network? If so, how to extract such subnetwork?" While empirical evidence has been accumulated about the existence of such unbiased subnetworks, these observations are mainly based on the guidance of ground-truth unbiased samples. Thus, it is unexplored how to discover the optimal subnetworks with biased training datasets in practice. To address this, here we first present our theoretical insight that alerts potential limitations of existing algorithms in exploring unbiased subnetworks in the presence of strong spurious correlations. We then further elucidate the importance of bias-conflicting samples on structure learning. Motivated by these observations, we propose a Debiased Contrastive Weight Pruning (DCWP) algorithm, which probes unbiased subnetworks without expensive group annotations. Experimental results demonstrate that our approach significantly outperforms state-of-the-art debiasing methods despite its considerable reduction in the number of parameters.

Recent studies on transfer learning have shown that selectively fine-tuning a subset of layers or customizing different learning rates for each layer can greatly improve robustness to out-of-distribution (OOD) data and retain generalization capability in the pre-trained models. However, most of these methods employ manually crafted heuristics or expensive hyper-parameter search, which prevent them from scaling up to large datasets and neural networks. To solve this problem, we propose Trainable Projected Gradient Method (TPGM) to automatically learn the constraint imposed for each layer for a fine-grained fine-tuning regularization. This is motivated by formulating fine-tuning as a bi-level constrained optimization problem. Specifically, TPGM maintains a set of projection radii, i.e., distance constraints between the fine-tuned model and the pre-trained model, for each layer, and enforces them through weight projections. To learn the constraints, we propose a bi-level optimization to automatically learn the best set of projection radii in an end-to-end manner. Theoretically, we show that the bi-level optimization formulation is the key to learn different constraints for each layer. Empirically, with little hyper-parameter search cost, TPGM outperforms existing fine-tuning methods in OOD performance while matching the best in-distribution (ID) performance. For example, when fine-tuned on DomainNet-Real and ImageNet, compared to vanilla fine-tuning, TPGM shows 22% and 10% relative OOD improvement respectively on their sketch counterparts.

We introduce Train/Test-Time Adaptation with Retrieval (T3AR), a method to adapt models both at train and test time by means of a retrieval module and a searchable pool of external samples. Before inference, T3AR adapts a given model to the downstream task using refined pseudo-labels and a self-supervised contrastive objective function whose noise distribution leverages retrieved real samples to improve feature adaptation on the target data manifold. The retrieval of real images is key to T3AR since it does not rely solely on synthetic data augmentations to compensate for the lack of adaptation data, as typically done by other adaptation algorithms. Furthermore, thanks to the retrieval module, our method gives the user or service provider the possibility to improve model adaptation on the downstream task by incorporating further relevant data or to fully remove samples that may no longer be available due to changes in user preference after deployment. First, we show that T3AR can be used at training time to improve downstream fine-grained classification over standard fine-tuning baselines, and the fewer the adaptation data the higher the relative improvement (up to 13%). Second, we apply T3AR for test-time adaptation and show that exploiting a pool of external images at test-time leads to more robust representations over existing methods on DomainNet-126 and VISDA-C, especially when few adaptation data are available (up to 8%).

We study the problem of how to train a "personalization-friendly" model such that given only the task descriptions, the model can be adapted to different end-users' needs, e.g., for accurately classifying different subsets of objects. One baseline approach is to train a "generic" model for classifying a wide range of objects, followed by class selection. In our experiments, we however found it suboptimal, perhaps because the model's weights are kept frozen without being personalized. To address this drawback, we propose Train-once-for-All PERsonalization (TAPER), a framework that is trained just once and can later customize a model for different end-users given their task descriptions. TAPER learns a set of "basis" models and a mixer predictor, such that given the task description, the weights (not the predictions!) of the basis models can be on the fly combined into a single "personalized" model. Via extensive experiments on multiple recognition tasks, we show that TAPER consistently outperforms the baseline methods in achieving a higher personalized accuracy. Moreover, we show that TAPER can synthesize a much smaller model to achieve comparable performance to a huge generic model, making it "deployment-friendly" to resource-limited end devices. Interestingly, even without end-users' task descriptions, TAPER can still be specialized to the deployed context based on its past predictions, making it even more "personalization-friendly".

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.

Tracking objects with persistence in cluttered and dynamic environments remains a difficult challenge for computer vision systems. In this paper, we introduce TCOW, a new benchmark and model for visual tracking through heavy occlusion and containment. We set up a task where the goal is to, given a video sequence, segment both the projected extent of the target object, as well as the surrounding container or occluder whenever one exists. To study this task, we create a mixture of synthetic and annotated real datasets to support both supervised learning and structured evaluation of model performance under various forms of task variation, such as moving or nested containment. We evaluate two recent transformer-based video models and find that while they can be surprisingly capable of tracking targets under certain settings of task variation, there remains a considerable performance gap before we can claim a tracking model to have acquired a true notion of object permanence.

Most existing multiple object tracking (MOT) methods that solely rely on appearance features struggle in tracking highly deformable objects. Other MOT methods that use motion clues to associate identities across frames have difficulty handling egocentric videos effectively or efficiently. In this work, we propose DETracker, a new MOT method that jointly detects and tracks deformable objects in egocentric videos. DETracker uses three novel modules, namely the motion disentanglement network (MDN), the patch association network (PAN) and the patch memory network (PMN), to explicitly tackle the difficulties caused by severe ego motion and fast morphing target objects. DETracker is end-to-end trainable and achieves near real-time speed. We also present DogThruGlasses, a large-scale deformable multi-object tracking dataset, with 150 videos and 73K annotated frames, collected by smart glasses. DETracker outperforms existing state-of-the-art method on the DogThruGlasses dataset and YouTube-Hand dataset.

We introduce a method for generating realistic pedestrian trajectories and full-body animations that can be controlled to meet user-defined goals. We draw on recent advances in guided diffusion modeling to achieve test-time controllability of trajectories, which is normally only associated with rule-based systems. Our guided diffusion model allows users to constrain trajectories through target waypoints, speed, and specified social groups while accounting for the surrounding environment context. This trajectory diffusion model is integrated with a novel physics-based humanoid controller to form a closed-loop, full-body pedestrian animation system capable of placing large crowds in a simulated environment with varying terrains. We further propose utilizing the value function learned during RL training of the animation controller to guide diffusion to produce trajectories better suited for particular scenarios such as collision avoidance and traversing uneven terrain.

Previous Knowledge Distillation based efficient image retrieval methods employ a lightweight network as the student model for fast inference. However, the lightweight student model lacks adequate representation capacity for effective knowledge imitation during the most critical early training period, causing final performance degeneration. To tackle this issue, we propose a Capacity Dynamic Distillation framework, which constructs a student model with editable representation capacity. Specifically, the employed student model is initially a heavy model to fruitfully learn distilled knowledge in the early training epochs, and the student model is gradually compressed during the training. To dynamically adjust the model capacity, our dynamic framework inserts a learnable convolutional layer within each residual block in the student model as the channel importance indicator. The indicator is optimized simultaneously by the image retrieval loss and the compression loss, and a retrieval-guided gradient resetting mechanism is proposed to release the gradient conflict. Extensive experiments show that our method has superior inference speed and accuracy, e.g., on the VeRi-776 dataset, given the ResNet101 as a teacher, our method saves 67.13% model parameters and 65.67% FLOPs without sacrificing accuracy.

In this paper, we study the problem of unsupervised object detection from 3D point clouds in self-driving scenes. We present a simple yet effective method that exploits (i) point clustering in near-range areas where the point clouds are dense, (ii) temporal consistency to filter out noisy unsupervised detections, (iii) translation equivariance of CNNs to extend the auto-labels to long range, and (iv) self-supervision for improving on its own. Our approach, OYSTER (Object Discovery via Spatio-Temporal Refinement), does not impose constraints on data collection (such as repeated traversals of the same location), is able to detect objects in a zero-shot manner without supervised finetuning (even in sparse, distant regions), and continues to self-improve given more rounds of iterative self-training. To better measure model performance in self-driving scenarios, we propose a new planning-centric perception metric based on distance-to-collision. We demonstrate that our unsupervised object detector significantly outperforms unsupervised baselines on PandaSet and Argoverse 2 Sensor dataset, showing promise that self-supervision combined with object priors can enable object discovery in the wild. For more information, visit the project website: https://waabi.ai/research/oyster.

With generative models proliferating at a rapid rate, there is a growing need for general purpose fake image detectors. In this work, we first show that the existing paradigm, which consists of training a deep network for real-vs-fake classification, fails to detect fake images from newer breeds of generative models when trained to detect GAN fake images. Upon analysis, we find that the resulting classifier is asymmetrically tuned to detect patterns that make an image fake. The real class becomes a 'sink' class holding anything that is not fake, including generated images from models not accessible during training. Building upon this discovery, we propose to perform real-vs-fake classification without learning; i.e., using a feature space not explicitly trained to distinguish real from fake images. We use nearest neighbor and linear probing as instantiations of this idea. When given access to the feature space of a large pretrained vision-language model, the very simple baseline of nearest neighbor classification has surprisingly good generalization ability in detecting fake images from a wide variety of generative models; e.g., it improves upon the SoTA by +15.07 mAP and +25.90% acc when tested on unseen diffusion and autoregressive models.

Sequence generation models have recently made significant progress in unifying various vision tasks. Although some auto-regressive models have demonstrated promising results in end-to-end text spotting, they use specific detection formats while ignoring various text shapes and are limited in the maximum number of text instances that can be detected. To overcome these limitations, we propose a UNIfied scene Text Spotter, called UNITS. Our model unifies various detection formats, including quadrilaterals and polygons, allowing it to detect text in arbitrary shapes. Additionally, we apply starting-point prompting to enable the model to extract texts from an arbitrary starting point, thereby extracting more texts beyond the number of instances it was trained on. Experimental results demonstrate that our method achieves competitive performance compared to state-of-the-art methods. Further analysis shows that UNITS can extract a larger number of texts than it was trained on. We provide the code for our method at https://github.com/clovaai/units.

Learning surface by neural implicit rendering has been a promising way for multi-view reconstruction in recent years. Existing neural surface reconstruction methods, such as NeuS and VolSDF, can produce reliable meshes from multi-view posed images. Although they build a bridge between volume rendering and Signed Distance Function (SDF), the accuracy is still limited. In this paper, we argue that this limited accuracy is due to the bias of their volume rendering strategies, especially when the viewing direction is close to be tangent to the surface. We revise and provide an additional condition for the unbiased volume rendering. Following this analysis, we propose a new rendering method by scaling the SDF field with the angle between the viewing direction and the surface normal vector. Experiments on simulated data indicate that our rendering method reduces the bias of SDF-based volume rendering. Moreover, there still exists non-negligible bias when the learnable standard deviation of SDF is large at early stage, which means that it is hard to supervise the rendered depth with depth priors. Alternatively we supervise zero-level set with surface points obtained from a pre-trained Multi-View Stereo network. We evaluate our method on the DTU dataset and show that it outperforms the state-of-the-arts neural implicit surface methods without mask supervision.

Deep neural networks have demonstrated promising performance on image recognition tasks. However, they may heavily rely on confounding factors, using irrelevant artifacts or bias within the dataset as the cue to improve performance. When a model performs decision-making based on these spurious correlations, it can become untrustable and lead to catastrophic outcomes when deployed in the real-world scene. In this paper, we explore and try to solve this problem in the context of skin cancer diagnosis. We introduce a human-in-the-loop framework in the model training process such that users can observe and correct the model's decision logic when confounding behaviors happen. Specifically, our method can automatically discover confounding factors by analyzing the co-occurrence behavior of the samples. It is capable of learning confounding concepts using easily obtained concept exemplars. By mapping the blackbox model's feature representation onto an explainable concept space, human users can interpret the concept and intervene via first order-logic instruction. We systematically evaluate our method on our newly crafted, well-controlled skin lesion dataset and several public skin lesion datasets. Experiments show that our method can effectively detect and remove confounding factors from datasets without any prior knowledge about the category distribution and does not require fully annotated concept labels. We also show that our method enables the model to focus on clinicalrelated concepts, improving the model's performance and trustworthiness during model inference.

Transferability of adversarial examples is critical for black-box deep learning model attacks. While most existing studies focus on enhancing the transferability of untargeted adversarial attacks, few of them studied how to generate transferable targeted adversarial examples that can mislead models into predicting a specific class. Moreover, existing transferable targeted adversarial attacks usually fail to sufficiently characterize the target class distribution, thus suffering from limited transferability. In this paper, we propose the Transferable Targeted Adversarial Attack (TTAA), which can capture the distribution information of the target class from both label-wise and feature-wise perspectives, to generate highly transferable targeted adversarial examples. To this end, we design a generative adversarial training framework consisting of a generator to produce targeted adversarial examples, and feature-label dual discriminators to distinguish the generated adversarial examples from the target class images. Specifically, we design the label discriminator to guide the adversarial examples to learn label-related distribution information about the target class. Meanwhile, we design a feature discriminator, which extracts the feature-wise information with strong cross-model consistency, to enable the adversarial examples to learn the transferable distribution information. Furthermore, we introduce the random perturbation dropping to further enhance the transferability by augmenting the diversity of adversarial examples used in the training process. Experiments demonstrate that our method achieves excellent performance on the transferability of targeted adversarial examples. The targeted fooling rate reaches 95.13% when transferred from VGG-19 to DenseNet-121, which significantly outperforms the state-of-the-art methods.

Towards stable human pose estimation from monocular images, there remain two main dilemmas. On the one hand, the different perspectives, i.e., front view, side view, and top view, appear the inconsistent performances due to the depth ambiguity. On the other hand, foot posture plays a significant role in complicated human pose estimation, i.e., dance and sports, and foot-ground interaction, but unfortunately, it is omitted in most general approaches and datasets. In this paper, we first propose the Cross-View Fusion (CVF) module to catch up with better 3D intermediate representation and alleviate the view inconsistency based on the vision transformer encoder. Then the optimization-based method is introduced to reconstruct the foot pose and foot-ground contact for the general multi-view datasets including AIST++ and Human3.6M. Besides, the reversible kinematic topology strategy is innovated to utilize the contact information into the full-body with foot pose regressor. Extensive experiments on the popular benchmarks demonstrate that our method outperforms the state-of-the-art approaches by achieving 40.1mm PA-MPJPE on the 3DPW test set and 43.8mm on the AIST++ test set.