RESC: REfine the SCore with Adaptive Transformer Head for End-to-end Object Detection
Wang Honglie,Jiang Rong,Xu Jian,Sun Shouqian
DOI: https://doi.org/10.1007/s00521-022-07089-5
2022-01-01
Neural Computing and Applications
Abstract:Most detection models employ many detection heads to output their prediction results independently. However, the locality of convolutional neural networks (CNN) causes the features extracted by adjacent convolution kernels to be very similar, which leads to duplicate prediction results. To tackle this issue, the hand-designed non-maximum suppression (NMS) procedure is proposed to remove the duplicate results. However, the NMS procedure cannot be applied to certain scenarios, such as the crowd scenarios, and requires careful adjustment of hyper-parameters. Therefore, end-to-end training is necessary to improve the detection ability on more scenarios. To this end, we propose a model that enables the network to adaptively identify duplicate objects and output non-repetitive results, which can effectively replace the hand-designed non-maximum suppression procedure. By adding differentiated priors to image features, and using Multi-Head Attention to enhance the global communication between features, our model can detect objects in an end-to-end manner. Our model can be easily applied to traditional one-stage detectors, e.g., FCOS and RetinaNet. While fast convergence and high recall rate are achieved, the accuracy is also significantly better than the baseline and outperforms many one-stage and two-stage methods. It also achieves the comparable performance as traditional detectors under the dense scene datasets CrowdHuman. Evaluation results demonstrate that our model with ResNet-50 can achieve 40.5% in $${\mathrm{AP}}$$ on COCO dataset and 89.2% in $${\mathrm{AP}}_{50}$$ on CrowdHuman dataset.