CVPR 2022 Papers

We address the problem of ensemble selection in transfer learning: Given a large pool of source models we want to select an ensemble of models which, after fine-tuning on the target training set, yields the best performance on the target test set. Since fine-tuning all possible ensembles is computationally prohibitive, we aim at predicting performance on the target dataset using a computationally efficient transferability metric. We propose several new transferability metrics designed for this task and evaluate them in a challenging and realistic transfer learning setup for semantic segmentation: we create a large and diverse pool of source models by considering 17 source datasets covering a wide variety of image domain, two different architectures, and two pre-training schemes. Given this pool, we then automatically select a subset to form an ensemble performing well on a given target dataset. We compare the ensemble selected by our method to two baselines which select a single source model, either (1) from the same pool as our method; or (2) from a pool containing large source models, each with similar capacity as an ensemble. Averaged over 17 target datasets, we outperform these baselines by 6.0% and 2.5% relative mean IoU, respectively.

Transfer learning has become a popular method for leveraging pre-trained models in computer vision. However, without performing computationally expensive fine-tuning, it is difficult to quantify which pre-trained source models are suitable for a specific target task, or, conversely, to which tasks a pre-trained source model can be easily adapted to. In this work, we propose Gaussian Bhattacharyya Coefficient (GBC), a novel method for quantifying transferability between a source model and a target dataset. In a first step we embed all target images in the feature space defined by the source model, and represent them with per-class Gaussians. Then, we estimate their pairwise class separability using the Bhattacharyya coefficient, yielding a simple and effective measure of how well the source model transfers to the target task. We evaluate GBC on image classification tasks in the context of dataset and architecture selection. Further, we also perform experiments on the more complex semantic segmentation transferability estimation task. We demonstrate that GBC outperforms state-of-the-art transferability metrics on most evaluation criteria in the semantic segmentation settings, matches the performance of top methods for dataset transferability in image classification, and performs best on architecture selection problems for image classification.

Removing adverse weather conditions like rain, fog, and snow from images is an important problem in many applications. Most methods proposed in the literature have been designed to deal with just removing one type of degradation. Recently, a CNN-based method using neural architecture search (All-in-One) was proposed to remove all the weather conditions at once. However, it has a large number of parameters as it uses multiple encoders to cater to each weather removal task and still has scope for improvement in its performance. In this work, we focus on developing an efficient solution for the all adverse weather removal problem. To this end, we propose TransWeather, a transformer-based end-to-end model with just a single encoder and a decoder that can restore an image degraded by any weather condition. Specifically, we utilize a novel transformer encoder using intra-patch transformer blocks to enhance attention inside the patches to effectively remove smaller weather degradations. We also introduce a transformer decoder with learnable weather type embeddings to adjust to the weather degradation at hand. TransWeather achieves significant improvements across multiple test datasets over both All-in-One network as well as methods fine-tuned for specific tasks. TransWeather is also validated on real world test images and found to be more effective than previous methods. Implementation code can be found in the supplementary document. Code is available at https://github.com/jeya-maria-jose/TransWeather.

Visual place recognition is a challenging task for applications such as autonomous driving navigation and mobile robot localization. Distracting elements presenting in complex scenes often lead to deviations in the perception of visual place. To address this problem, it is crucial to integrate information from only task-relevant regions into image representations. In this paper, we introduce a novel holistic place recognition model, TransVPR, based on vision Transformers. It benefits from the desirable property of the self-attention operation in Transformers which can naturally aggregate task-relevant features. Attentions from multiple levels of the Transformer, which focus on different regions of interest, are further combined to generate a global image representation. In addition, the output tokens from Transformer layers filtered by the fused attention mask are considered as key-patch descriptors, which are used to perform spatial matching to re-rank the candidates retrieved by the global image features. The whole model allows end-to-end training with a single objective and image-level supervision. TransVPR achieves state-of-the-art performance on several real-world benchmarks while maintaining low computational time and storage requirements.

Recognizing transformation types applied to a video clip (RecogTrans) is a long-established paradigm for self-supervised video representation learning, which achieves much inferior performance compared to instance discrimination approaches (InstDisc) in recent works. However, based on a thorough comparison of representative RecogTrans and InstDisc methods, we observe the great potential of RecogTrans on both semantic-related and temporal-related downstream tasks. Based on hard-label classification, existing RecogTrans approaches suffer from noisy supervision signals in pre-training. To mitigate this problem, we developed TransRank, a unified framework for recognizing Transformations in a Ranking formulation. TransRank provides accurate supervision signals by recognizing transformations relatively, consistently outperforming the classification-based formulation. Meanwhile, the unified framework can be instantiated with an arbitrary set of temporal or spatial transformations, demonstrating good generality. With a ranking-based formulation and several empirical practices, we achieve competitive performance on video retrieval and action recognition.Under the same setting, TransRank surpasses the previous state-of-the-art method by 6.4% on UCF101 and 8.3% on HMDB51 for action recognition (Top1 Acc); improves video retrieval on UCF101 by 20.4% (R@1). The promising results validate that RecogTrans is still a worth exploring paradigm for video self-supervised learning. Codes will be released at https://github.com/kennymckormick/TransRank.

Counting repetitive actions are widely seen in human activities such as physical exercise. Existing methods focus on performing repetitive action counting in short videos, which is tough for dealing with longer videos in more realistic scenarios. In the data-driven era, the degradation of such generalization capability is mainly attributed to the lack of long video datasets. To complement this margin, we introduce a new large-scale repetitive action counting dataset covering a wide variety of video lengths, along with more realistic situations where action interruption or action inconsistencies occur in the video. Besides, we also provide a fine-grained annotation of the action cycles instead of just counting annotation along with a numerical value. Such a dataset contains 1,451 videos with about 20,000 annotations, which is more challenging. For repetitive action counting towards more realistic scenarios, we further propose encoding multi-scale temporal correlation with transformers that can take into account both performance and efficiency. Furthermore, with the help of fine-grained annotation of action cycles, we propose a density map regression-based method to predict the action period, which yields better performance with sufficient interpretability. Our proposed method outperforms state-of-the-art methods on all datasets and also achieves better performance on the unseen dataset without fine-tuning. The dataset and code are available.

Mixup-based augmentation has been found to be effective for generalizing models during training, especially for Vision Transformers (ViTs) since they can easily overfit. However, previous mixup-based methods have an underlying prior knowledge that the linearly interpolated ratio of targets should be kept the same as the ratio proposed in input interpolation. This may lead to a strange phenomenon that sometimes there is no valid object in the mixed image due to the random process in augmentation but there is still response in the label space. To bridge such gap between the input and label spaces, we propose TransMix, which mixes labels based on the attention maps of Vision Transformers. The confidence of the label will be larger if the corresponding input image is weighted higher by the attention map. TransMix is embarrassingly simple and can be implemented in just a few lines of code without introducing any extra parameters and FLOPs to ViT-based models. Experimental results show that our method can consistently improve various ViT-based models at scales on ImageNet classification. After pre-trained with TransMix on ImageNet, the ViT-based models also demonstrate better transferability to semantic segmentation, object detection and instance segmentation. TransMix also exhibits to be more robust when evaluating on 4 different benchmarks. Code is publicly available at https://github.com/Beckschen/TransMix.

The dominant CNN-based methods for cross-view image geo-localization rely on polar transform and fail to model global correlation. We propose a pure transformer-based approach (TransGeo) to address these limitations from a different perspective. TransGeo takes full advantage of the strengths of transformer related to global information modeling and explicit position information encoding. We further leverage the flexibility of transformer input and propose an attention-guided non-uniform cropping method, so that uninformative image patches are removed with negligible drop on performance to reduce computation cost. The saved computation can be reallocated to increase resolution only for informative patches, resulting in performance improvement with no additional computation cost. This "attend and zoom-in" strategy is highly similar to human behavior when observing images. Remarkably, TransGeo achieves state-of-the-art results on both urban and rural datasets, with significantly less computation cost than CNN-based methods. It does not rely on polar transform and infers faster than CNN-based methods. Code is available at https://github.com/Jeff-Zilence/TransGeo2022.

LiDAR and camera are two important sensors for 3D object detection in autonomous driving. Despite the increasing popularity of sensor fusion in this field, the robustness against inferior image conditions, e.g., bad illumination and sensor misalignment, is under-explored. Existing fusion methods are easily affected by such conditions, mainly due to a hard association of LiDAR points and image pixels, established by calibration matrices. We propose TransFusion, a robust solution to LiDAR-camera fusion with a soft-association mechanism to handle inferior image conditions. Specifically, our TransFusion consists of convolutional backbones and a detection head based on a transformer decoder. The first layer of the decoder predicts initial bounding boxes from a LiDAR point cloud using a sparse set of object queries, and its second decoder layer adaptively fuses the object queries with useful image features, leveraging both spatial and contextual relationships. The attention mechanism of the transformer enables our model to adaptively determine where and what information should be taken from the image, leading to a robust and effective fusion strategy. We additionally design an image-guided query initialization strategy to deal with objects that are difficult to detect in point clouds. TransFusion achieves state-of-the-art performance on large-scale datasets. We provide extensive experiments to demonstrate its robustness against degenerated image quality and calibration errors. We also extend the proposed method to the 3D tracking task and achieve the 1st place in the leaderboard of nuScenes tracking, showing its effectiveness and generalization capability.

Recent advances like StyleGAN have promoted the growth of controllable facial editing. To address its core challenge of attribute decoupling in a single latent space, attempts have been made to adopt dual-space GAN for better disentanglement of style and content representations. Nonetheless, these methods are still incompetent to obtain plausible editing results with high controllability, especially for complicated attributes. In this study, we highlight the importance of interaction in a dual-space GAN for more controllable editing. We propose TransEditor, a novel Transformer-based framework to enhance such interaction. Besides, we develop a new dual-space editing and inversion strategy to provide additional editing flexibility. Extensive experiments demonstrate the superiority of the proposed framework in image quality and editing capability, suggesting the effectiveness of TransEditor for highly controllable facial editing. Code and models are publicly available at https://github.com/BillyXYB/TransEditor.

We focus on the task of estimating a physically plausible articulated human motion from monocular video. Existing approaches that do not consider physics often produce temporally inconsistent output with motion artifacts, while state-of-the-art physics-based approaches have either been shown to work only in controlled laboratory conditions or consider simplified body-ground contact limited to feet. This paper explores how these shortcomings can be addressed by directly incorporating a fully-featured physics engine into the pose estimation process. Given an uncontrolled, real-world scene as input, our approach estimates the ground-plane location and the dimensions of the physical body model. It then recovers the physical motion by performing trajectory optimization. The advantage of our formulation is that it readily generalizes to a variety of scenes that might have diverse ground properties and supports any form of self-contact and contact between the articulated body and scene geometry. We show that our approach achieves competitive results with respect to existing physics-based methods on the Human3.6M benchmark, while being directly applicable without re-training to more complex dynamic motions from the AIST benchmark and to uncontrolled internet videos.

Recently, Vision Transformer (ViT) has achieved remarkable success in several computer vision tasks. The progresses are highly relevant to the architecture design, then it is worthwhile to propose Transformer Architecture Search (TAS) to search for better ViTs automatically. However, current TAS methods are time-consuming and existing zero-cost proxies in CNN do not generalize well to the ViT search space according to our experimental observations. In this paper, for the first time, we investigate how to conduct TAS in a training-free manner and devise an effective training-free TAS (TF-TAS) scheme. Firstly, we observe that the properties of multi-head self-attention (MSA) and multi-layer perceptron (MLP) in ViTs are quite different and that the synaptic diversity of MSA affects the performance notably. Secondly, based on the observation, we devise a modular strategy in TF-TAS that evaluates and ranks ViT architectures from two theoretical perspectives: synaptic diversity and synaptic saliency, termed as DSS-indicator. With DSS-indicator, evaluation results are strongly correlated with the test accuracies of ViT models. Experimental results demonstrate that our TF-TAS achieves a competitive performance against the state-of-the-art manually or automatically design ViT architectures, and it promotes the searching efficiency in ViT search space greatly: from about 24 GPU days to less than 0.5 GPU days. Moreover, the proposed DSS-indicator outperforms the existing cutting-edge zero-cost approaches (e.g., TE-score and NASWOT).

Training quantised neural networks (QNNs) is a non-differentiable optimisation problem since weights and features are output by piecewise constant functions. The standard solution is to apply the straight-through estimator (STE), using different functions during the inference and gradient computation steps. Several STE variants have been proposed in the literature aiming to maximise the task accuracy of the trained network. In this paper, we analyse STE variants and study their impact on QNN training. We first observe that most such variants can be modelled as stochastic regularisations of stair functions; although this intuitive interpretation is not new, our rigorous discussion generalises to further variants. Then, we analyse QNNs mixing different regularisations, finding that some suitably synchronised smoothing of each layer map is required to guarantee pointwise compositional convergence to the target discontinuous function. Based on these theoretical insights, we propose additive noise annealing (ANA), a new algorithm to train QNNs encompassing standard STE and its variants as special cases. When testing ANA on the CIFAR-10 image classification benchmark, we find that the major impact on task accuracy is not due to the qualitative shape of the regularisations but to the proper synchronisation of the different STE variants used in a network, in accordance with the theoretical results.

Modeling in computer vision has long been dominated by convolutional neural networks (CNNs). Recently, in light of the excellent performances of self-attention mechanism in the language field, transformers tailored for visual data have drawn numerous attention and triumphed CNNs in various vision tasks. These vision transformers heavily rely on large-scale pre-training to achieve competitive accuracy, which not only hinders the freedom of architectural design in downstream tasks like object detection, but also causes learning bias and domain mismatch in the fine-tuning stages. To this end, we aim to get rid of the "pre-train & fine-tune" paradigm of vision transformer and train transformer based object detector from scratch. Some earlier work in the CNNs era have successfully trained CNNs based detectors without pre-training, unfortunately, their findings do not generalize well when the backbone is switched from CNNs to vision transformer. Instead of proposing a specific vision transformer based detector, in this work, our goal is to reveal the insights of training vision transformer based detectors from scratch. In particular, we expect those insights can help other researchers and practitioners, and inspire more interesting research in other fields, such as semantic segmentation, visual-linguistic pre-training, etc. One of the key findings is that both architectural changes and more epochs play critical roles in training vision transformer based detectors from scratch. Experiments on MS COCO datasets demonstrate that vision transformer based detectors trained from scratch can also achieve similar performances to their counterparts with ImageNet pre-training.

Spiking Neural Network (SNN) is a promising energy-efficient AI model when implemented on neuromorphic hardware. However, it is a challenge to efficiently train SNNs due to their non-differentiability. Most existing methods either suffer from high latency (i.e., long simulation time steps), or cannot achieve as high performance as Artificial Neural Networks (ANNs). In this paper, we propose the Differentiation on Spike Representation (DSR) method, which could achieve high performance that is competitive to ANNs yet with low latency. First, we encode the spike trains into spike representation using (weighted) firing rate coding. Based on the spike representation, we systematically derive that the spiking dynamics with common neural models can be represented as some sub-differentiable mapping. With this viewpoint, our proposed DSR method trains SNNs through gradients of the mapping and avoids the common non-differentiability problem in SNN training. Then we analyze the error when representing the specific mapping with the forward computation of the SNN. To reduce such error, we propose to train the spike threshold in each layer, and to introduce a new hyperparameter for the neural models. With these components, the DSR method can achieve state-of-the-art SNN performance with low latency on both static and neuromorphic datasets, including CIFAR-10, CIFAR-100, ImageNet, and DVS-CIFAR10.

We present an approach for tracking people in monocular videos by predicting their future 3D representations. To achieve this, we first lift people to 3D from a single frame in a robust manner. This lifting includes information about the 3D pose of the person, their location in the 3D space, and the 3D appearance. As we track a person, we collect 3D observations over time in a tracklet representation. Given the 3D nature of our observations, we build temporal models for each one of the previous attributes. We use these models to predict the future state of the tracklet, including 3D appearance, 3D location, and 3D pose. For a future frame, we compute the similarity between the predicted state of a tracklet and the single frame observations in a probabilistic manner. Association is solved with simple Hungarian matching, and the matches are used to update the respective tracklets. We evaluate our approach on various benchmarks and report state-of-the-art results. Code and models are available at: https://brjathu.github.io/PHALP.

The challenging task of multi-object tracking (MOT) requires simultaneous reasoning about track initialization, identity, and spatio-temporal trajectories. We formulate this task as a frame-to-frame set prediction problem and introduce TrackFormer, an end-to-end trainable MOT approach based on an encoder-decoder Transformer architecture. Our model achieves data association between frames via attention by evolving a set of track predictions through a video sequence. The Transformer decoder initializes new tracks from static object queries and autoregressively follows existing tracks in space and time with the conceptually new and identity preserving track queries. Both query types benefit from self- and encoder-decoder attention on global frame-level features, thereby omitting any additional graph optimization or modeling of motion and/or appearance. TrackFormer introduces a new tracking-by-attention paradigm and while simple in its design is able to achieve state-of-the-art performance on the task of multi-object tracking (MOT17) and segmentation (MOTS20). The code is available at https://github.com/timmeinhardt/trackformer

Optical flow, which captures motion information across frames, is exploited in recent video inpainting methods through propagating pixels along its trajectories. However, the hand-crafted flow-based processes in these methods are applied separately to form the whole inpainting pipeline. Thus, they are less efficient and rely heavily on the intermediate results from earlier stages. In this paper, we propose an End-to-End framework for Flow-Guided Video Inpainting through elaborately designed three trainable modules, namely, flow completion, feature propagation, and content hallucination modules. The three modules correspond with the three stages of previous flow-based methods but can be jointly optimized, leading to a more efficient and effective inpainting process. Experimental results demonstrate that our proposed method outperforms state-of-the-art methods both qualitatively and quantitatively and shows promising efficiency. The code is available at https://github.com/MCG-NKU/E2FGVI.