Search

Class activation maps (CAMs) have been widely used on weakly-supervised object localization, which generate attention maps for specific categories in an image. Since CAMs can be obtained using category annotation, which is included in the annotation information of fully-supervised object detection. Therefore, how to adopt attention information in CAMs to improve the performance of fully-supervised object detection is an interesting problem. In this paper, we propose CAM R-CNN to deal with object detection, in which the category-aware attention maps provided by CAMs are integrated into the process of object detection. CAM R-CNN follows the common pipeline of the recent query-based object detectors in an end-to-end fashion, while two key CAM modules are embedded into the process. Specifically, E-CAM module provides embedding-level attention via fusing proposal features and attention information in CAMs with a transformer encoder, and S-CAM module supplies spatial-level attention by multiplying feature maps with the top-activated attention map provided by CAMs. In our experiments, CAM R-CNN demonstrates its superiority compared to several strong baselines on the challenging COCO dataset. Furthermore, we show that S-CAM module can be applied to two-stage detectors such as Faster R-CNN and Cascade R-CNN with consistent gains. [ABSTRACT FROM AUTHOR]

We attempt to reduce the computational costs in vision transformers (ViTs), which increase quadratically in the token number. We present a novel training paradigm that trains only one ViT model at a time, but is capable of providing improved image recognition performance with various computational costs. Here, the trained ViT model, termed super vision transformer (SuperViT), is empowered with the versatile ability to solve incoming patches of multiple sizes as well as preserve informative tokens with multiple keeping rates (the ratio of keeping tokens) to achieve good hardware efficiency for inference, given that the available hardware resources often change from time to time. Experimental results on ImageNet demonstrate that our SuperViT can considerably reduce the computational costs of ViT models with even performance increase. For example, we reduce 2 × FLOPs of DeiT-S while increasing the Top-1 accuracy by 0.2% and 0.7% for 1.5 × reduction. Also, our SuperViT significantly outperforms existing studies on efficient vision transformers. For example, when consuming the same amount of FLOPs, our SuperViT surpasses the recent state-of-the-art EViT by 1.1% when using DeiT-S as their backbones. The project of this work is made publicly available at https://github.com/lmbxmu/SuperViT. [ABSTRACT FROM AUTHOR]