Richer Information Transformer for Object Detection

While Convolutional Neural Networks (CNNs) have been dominated computer vision tasks such as object detection and instance segmentation for a long time, recently Vision Transformers (ViTs) are showing promising performance in these tasks nowadays. Though CNNs can efficiently decrease local redundancy by convolution within a small neighborhood, the limited receptive field makes it hard to capture global dependency. Alternatively, ViTs can effectively capture long-range dependency via self-attention mechanism which also produce quadratic computation complexity to the image size input. In this paper, we propose a module composed of several groups of convs and activation functions to make up for the lack of comprehensive information in ViTs for extracting features, so that conv and transformer can achieve a complementary advantage. We also introduce channel attention module to capture the channel information, which arising from the frequent manipulations to channels during the calculation process of self-attention. In the absence of pretrained data, our model achieves 40.3 box AP and 37.1 mask AP on COCO object detection task, surpassing state-of-art Swin Transformers backbone respectively by +8.8, +6.7 respectively under the similar FLOPs settings.