Ying Li

Shihao Wang, Xiaohui Jiang, Ying Li

The dominant multi-camera 3D detection paradigm is based on explicit 3D feature construction, which requires complicated indexing of local image-view features via 3D-to-2D projection. Other methods implicitly introduce geometric positional encoding and perform global attention (e.g., PETR) to build the relationship between image tokens and 3D objects. The 3D-to-2D perspective inconsistency and global attention lead to a weak correlation between foreground tokens and queries, resulting in slow convergence. We propose Focal-PETR with instance-guided supervision and spatial alignment module to adaptively focus object queries on discriminative foreground regions. Focal-PETR additionally introduces a down-sampling strategy to reduce the consumption of global attention. Due to the highly parallelized implementation and down-sampling strategy, our model, without depth supervision, achieves leading performance on the large-scale nuScenes benchmark and a superior speed of 30 FPS on a single RTX3090 GPU. Extensive experiments show that our method outperforms PETR while consuming 3x fewer training hours. The code will be made publicly available.

Zinan Xiong, Chenxi Wang, Ying Li et al.

Convolutional neural networks (CNNs) have been widely utilized in many computer vision tasks. However, CNNs have a fixed reception field and lack the ability of long-range perception, which is crucial to human pose estimation. Due to its capability to capture long-range dependencies between pixels, transformer architecture has been adopted to computer vision applications recently and is proven to be a highly effective architecture. We are interested in exploring its capability in human pose estimation, and thus propose a novel model based on transformer architecture, enhanced with a feature pyramid fusion structure. More specifically, we use pre-trained Swin Transformer as our backbone and extract features from input images, we leverage a feature pyramid structure to extract feature maps from different stages. By fusing the features together, our model predicts the keypoint heatmap. The experiment results of our study have demonstrated that the proposed transformer-based model can achieve better performance compared to the state-of-the-art CNN-based models.

Xizhe Xue, Ying Li, Xiaoyue Yin et al.

Discriminative correlation filters (DCF) and siamese networks have achieved promising performance on visual tracking tasks thanks to their superior computational efficiency and reliable similarity metric learning, respectively. However, how to effectively take advantages of powerful deep networks, while maintaining the real-time response of DCF, remains a challenging problem. Embedding the cross-correlation operator as a separate layer into siamese networks is a popular choice to enhance the tracking accuracy. Being a key component of such a network, the correlation layer is updated online together with other parts of the network. Yet, when facing serious disturbance, fused trackers may still drift away from the target completely due to accumulated errors. To address these issues, we propose a coarse-to-fine tracking framework, which roughly infers the target state via an online-updating DCF module first and subsequently, finely locates the target through an offline-training asymmetric siamese network (ASN). Benefitting from the guidance of DCF and the learned channel weights obtained through exploiting the given ground-truth template, ASN refines feature representation and implements precise target localization. Systematic experiments on five popular tracking datasets demonstrate that the proposed DCF-ASN achieves the state-of-the-art performance while exhibiting good tracking efficiency.