Bird's-Eye-View (BEV) representation has emerged as a mainstream paradigm for multi-view 3D object detection, demonstrating impressive perceptual capabilities. However, existing methods overlook the geometric quality of BEV representation, leaving it in a low-resolution state and failing to restore the authentic geometric information of the scene. In this paper, we identify the drawbacks of previous approaches that limit the geometric quality of BEV representation and propose Radial-Cartesian BEV Sampling (RC-Sampling), which outperforms other feature transformation methods in efficiently generating high-resolution dense BEV representation to restore fine-grained geometric information. Additionally, we design a novel In-Box Label to substitute the traditional depth label generated from the LiDAR points. This label reflects the actual geometric structure of objects rather than just their surfaces, injecting real-world geometric information into the BEV representation. In conjunction with the In-Box Label, Centroid-Aware Inner Loss (CAI Loss) is developed to capture the inner geometric structure of objects. Finally, we integrate the aforementioned modules into a novel multi-view 3D object detector, dubbed GeoBEV, which achieves a state-of-the-art result of 66.2% NDS on the nuScenes test set.

Despite its significant success, object detection in traffic and transportation scenarios requires time-consuming and laborious efforts in acquiring high-quality labeled data. Therefore, Unsupervised Domain Adaptation (UDA) for object detection has recently gained increasing research attention. UDA for object detection has been dominated by domain alignment methods, which achieve top performance. Recently, self-labeling methods have gained popularity due to their simplicity and efficiency. In this paper, we investigate the limitations that prevent self-labeling detectors from achieving commensurate performance with domain alignment methods. Specifically, we identify the high proportion of simple samples during training, i.e., the simple-label bias, as the central cause. We propose a novel approach called De-Simplifying Pseudo Labels (DeSimPL) to mitigate the issue. DeSimPL utilizes an instance-level memory bank to implement an innovative pseudo label updating strategy. Then, adversarial samples are introduced during training to enhance the proportion. Furthermore, we propose an adaptive weighted loss to avoid the model suffering from an abundance of false positive pseudo labels in the late training period. Experimental results demonstrate that DeSimPL effectively reduces the proportion of simple samples during training, leading to a significant performance improvement for self-labeling detectors. Extensive experiments conducted on four benchmarks validate our analysis and conclusions.

Recent attention has been devoted to the pursuit of learning semantic segmentation models exclusively from image tags, a paradigm known as image-level Weakly Supervised Semantic Segmentation (WSSS). Existing attempts adopt the Class Activation Maps (CAMs) as priors to mine object regions yet observe the imbalanced activation issue, where only the most discriminative object parts are located. In this paper, we argue that the distribution discrepancy between the discriminative and the non-discriminative parts of objects prevents the model from producing complete and precise pseudo masks as ground truths. For this purpose, we propose a Pixel-Level Domain Adaptation (PLDA) method to encourage the model in learning pixel-wise domain-invariant features. Specifically, a multi-head domain classifier trained adversarially with the feature extraction is introduced to promote the emergence of pixel features that are invariant with respect to the shift between the source (i.e., the discriminative object parts) and the target (i.e., the non-discriminative object parts) domains. In addition, we come up with a Confident Pseudo-Supervision strategy to guarantee the discriminative ability of each pixel for the segmentation task, which serves as a complement to the intra-image domain adversarial training. Our method is conceptually simple, intuitive and can be easily integrated into existing WSSS methods. Taking several strong baseline models as instances, we experimentally demonstrate the effectiveness of our approach under a wide range of settings.

Human pose estimation and tracking are fundamental tasks for understanding human behaviors in videos. Existing top-down framework-based methods usually perform three-stage tasks: human detection, pose estimation and tracking. Although promising results have been achieved, these methods rely heavily on high-performance detectors and may fail to track persons who are occluded or miss-detected. To overcome these problems, in this paper, we develop a novel keypoint confidence network and a tracking pipeline to improve human detection and pose estimation in top-down approaches. Specifically, the keypoint confidence network is designed to determine whether each keypoint is occluded, and it is incorporated into the pose estimation module. In the tracking pipeline, we propose the Bboxrevision module to reduce missing detection and the ID-retrieve module to correct lost trajectories, improving the performance of the detection stage. Experimental results show that our approach is universal in human detection and pose estimation, achieving state-of-the-art performance on both PoseTrack 2017 and 2018 datasets.

Although much progress has been made recently in 3D face reconstruction, most previous work has been devoted to predicting accurate and fine-grained 3D shapes. In contrast, relatively little work has focused on generating high-fidelity face textures. Compared with the prosperity of photo-realistic 2D face image generation, high-fidelity 3D face texture generation has yet to be studied. In this paper, we propose a novel UV map generation model that predicts the UV map from a single face image. The model consists of a UV sampler and a UV generator. By selectively sampling the input face image's pixels and adjusting their relative locations, the UV sampler generates an incomplete UV map that could faithfully reconstruct the original face. Missing textures in the incomplete UV map are further full-filled by the UV generator. The training is based on pseudo ground truth blended by the 3DMM texture and the input face texture, thus weakly supervised. To deal with the artifacts in the imperfect pseudo UV map, multiple UV map and face image discriminators are leveraged.

Occlusions are very common in face images in the wild, leading to the degraded performance of face-related tasks. Although much effort has been devoted to removing occlusions from face images, the varying shapes and textures of occlusions still challenge the robustness of current methods. As a result, current methods either rely on manual occlusion masks or only apply to specific occlusions. This paper proposes a novel face de-occlusion model based on face segmentation and 3D face reconstruction, which is robust to arbitrary kinds of face occlusions. The proposed model consists of a 3D face reconstruction module, a face segmentation module, and an image generation module. With the face prior and the occlusion mask predicted by the first two, respectively, the image generation module can faithfully recover the missing facial textures. To supervise the training, we further build a large occlusion dataset, with both manually labeled and synthetic occlusions. Qualitative and quantitative results demonstrate the effectiveness and robustness of the proposed method.

This work addresses the task of action recognition in video sequences. In real world applications, this task is quite challenging due to the complex background of video content, the similarities between different types of actions, the dependence on a large amount of annotated data, and so on. Most of the existing methods fail to distinguish similar actions with the same static appearance and motion pattern. We attempt to address this issue from the perspective of a local-global view, considering videos as combinations of a set of action units (local semantic information) and their relations along temporal dimension (global relation information). To achieve this end, we propose a novel Local-global Networks (LgNet) to enhance recognition of similar action. Besides, we propose an end-to-end training method to decrease the reliance on annotated data. It combines self-supervised learning and supervised learning, which not only enables the model to learn video representations from a large number unannotated data but also avoids subsequent finetuning. The proposed training method can be flexibly equipped to a wide array of vision tasks. Experiments on several benchmark datasets show that our proposed model and training method achieve state-of-the-art performance.

Though image-level weakly supervised semantic seg-mentation (WSSS) has achieved great progress with Class Activation Maps (CAMs) as the cornerstone, the large su-pervision gap between classification and segmentation still hampers the model to generate more complete and precise pseudo masks for segmentation. In this study, we propose weakly-supervised pixel-to-prototype contrast that can provide pixel-level supervisory signals to narrow the gap. Guided by two intuitive priors, our method is executed across different views and within per single view of an image, aiming to impose cross-view feature semantic consistency regularization and facilitate intra(inter)-class compactness(dispersion) of the feature space. Our method can be seamlessly incorporated into existing WSSS models with-out any changes to the base networks and does not incur any extra inference burden. Extensive experiments manifest that our method consistently improves two strong baselines by large margins, demonstrating the effectiveness. Specifically, built on top of SEAM, we improve the initial seed mIoU on PASCAL VOC 2012 from 55.4% to 61.5%. Moreover, armed with our method, we increase the segmentation mIoU of EPS from 70.8% to 73.6%, achieving new state-of-the-art.

In this paper, we propose a transformer based approach for visual grounding. Unlike existing proposal-and-rank frameworks that rely heavily on pretrained object detectors or proposal-free frameworks that upgrade an off-the-shelf one-stage detector by fusing textual embeddings, our approach is built on top of a transformer encoder-decoder and is independent of any pretrained detectors or word embedding models. Termed as VGTR – Visual Grounding with TRansformers, our approach is designed to learn semantic-discriminative visual features under the guidance of the textual description without harming their location ability. This information flow enables our VGTR to have a strong capability in capturing context-level semantics of both vision and language modalities, rendering us to aggregate accurate visual clues implied by the description to locate the interested object instance. Experiments show that our method outperforms state-of-the-art proposal-free approaches by a considerable margin on four benchmarks.

Transformers have been successfully applied to the visual tracking task and significantly promote tracking performance. The self-attention mechanism designed to model long-range dependencies is the key to the success of Transformers. However, self-attention lacks focusing on the most relevant information in the search regions, making it easy to be distracted by background. In this paper, we relieve this issue with a sparse attention mechanism by focusing the most relevant information in the search regions, which enables a much accurate tracking. Furthermore, we introduce a double-head predictor to boost the accuracy of foreground-background classification and regression of target bounding boxes, which further improve the tracking performance. Extensive experiments show that, without bells and whistles, our method significantly outperforms the state-of-the-art approaches on LaSOT, GOT-10k, TrackingNet, and UAV123, while running at 40 FPS. Notably, the training time of our method is reduced by 75% compared to that of TransT. The source code and models are available at https://github.com/fzh0917/SparseTT.

Modeling relations between actors is critical for understanding group activities of dynamic scenes. Existing Group Activity Recognition (GAR) methods usually build strong connection in each actor pair. However, not all the connetions are necessary because not all actors are visible or related to each other. Based on this observation, we provide a Sparse Relation Graph (SRG) for GAR, in which the key relations are focused to mine more discriminative features. Then a graph convolutional network is designed for automatically learning the key relations. Extensive experiments on two popular group activity datasets, the Volleyball dataset and the Collective Activity dataset, demonstrate the effectiveness of our method. Especially in the Volleyball dataset, SRG can get better performance with less but delicate information.

Self-training has shown great potential in semi-supervised learning. Its core idea is to use the model learned on labeled data to generate pseudo-labels for unlabeled samples, and in turn teach itself. To obtain valid supervision, active attempts typically employ a momentum teacher for pseudo-label prediction yet observe the confirmation bias issue, where the incorrect predictions may provide wrong supervision signals and get accumulated in the training process. The primary cause of such a drawback is that the prevailing self-training framework acts as guiding the current state with previous knowledge because the teacher is updated with the past student only. To alleviate this problem, we propose a novel self-training strategy, which allows the model to learn from the future. Concretely, at each training step, we first virtually optimize the student (i.e., caching the gradients without applying them to the model weights), then update the teacher with the virtual future student, and finally ask the teacher to produce pseudo-labels for the current student as the guidance. In this way, we manage to improve the quality of pseudo-labels and thus boost the performance. We also develop two variants of our future-self-training (FST) framework through peeping at the future both deeply (FST-D) and widely (FST-W). Taking the tasks of unsupervised domain adaptive semantic segmentation and semi-supervised semantic segmentation as the instances, we experimentally demonstrate the effectiveness and superiority of our approach under a wide range of settings. Code is available at https://github.com/usr922/FST.

Boosting performance of the offline trained siamese trackers is getting harder nowadays since the fixed information of the template cropped from the first frame has been almost thoroughly mined, but they are poorly capable of resisting target appearance changes. Existing trackers with template updating mechanisms rely on time-consuming numerical optimization and complex hand-designed strategies to achieve competitive performance, hindering them from real-time tracking and practical applications. In this paper, we propose a novel tracking framework built on top of a space-time memory network that is competent to make full use of historical information related to the target for better adapting to appearance variations during tracking. Specifically, a novel memory mechanism is introduced, which stores the historical information of the target to guide the tracker to focus on the most informative regions in the current frame. Furthermore, the pixel-level similarity computation of the memory network enables our tracker to generate much more accurate bounding boxes of the target. Extensive experiments and comparisons with many competitive trackers on challenging large-scale benchmarks, OTB-2015, TrackingNet, GOT-10k, LaSOT, UAV123, and VOT2018, show that, without bells and whistles, our tracker outperforms all previous state-of-the-art real-time methods while running at 37 FPS. The code is available at https: //github.com/fzh0917/STMTrack.

The existing auto-encoder based face pose editing methods primarily focus on modeling the identity preserving ability during pose synthesis, but are less able to preserve the image style properly, which refers to the color, brightness, saturation, etc. In this paper, we take advantage of the well-known frontal/profile optical illusion and present a novel two-stage approach to solve the aforementioned dilemma, where the task of face pose manipulation is cast into face inpainting. By selectively sampling pixels from the input face and slightly adjust their relative locations with the proposed “Pixel Attention Sampling” module, the face editing result faithfully keeps the identity information as well as the image style unchanged. By leveraging high-dimensional embedding at the inpainting stage, finer details are generated. Further, with the 3D facial landmarks as guidance, our method is able to manipulate face pose in three degrees of freedom, i.e., yaw, pitch, and roll, resulting in more flexible face pose editing than merely controlling the yaw angle as usually achieved by the current state-of-the-art. Both the qualitative and quantitative evaluations validate the superiority of the proposed approach.

In stereo matching, the correctness of stereo pairs matches, also called confidence, is used to improve the dense disparity estimation result. In this paper, we propose a multi-modal deep learning approach for stereo matching confidence estimation. The input of our method is composed of two modalities, the initial disparity maps, and its reference color image. To effectively combine these two modalities, we explore and study multiple convolutional neural network (CNN) structures for our specific confidence prediction tasks. To the best of our knowledge, this is the first approach reported in the literature combining multiple modalities and patch based deep learning to predict the confidence. The experiments on KITTI datasets demonstrate that our multi-modal confidence network can significantly outperform the state-of-the-art methods.