Improving the Detection Performance of Sparse R-CNN with Different Necks

Abstract

Sparse R-CNN uses a purely sparse method to detect objects and achieves good results. However, it does not make full use of the features extracted from the image, so its detection performance needs to be further improved. And we propose Sparse R-CNNv1 and Sparse R-CNNv2. In these algorithms, we use VOVNet with attention mechanism to replace ResNet of the original Sparse R-CNN as our backbone. In addition, we also use two different improved neck networks, Augpan and FPNencoder, to further improve the detection performance of the algorithm from the perspective of feature fusion and increasing the receptive field of each layer, respectively. Our algorithms are trained and verified on COCO2017, and the experimental results show that Sparser-CNNv1 achieves 45.0 AP and Sparser-CNNV2 achieves 45.3 AP, higher than the original SparseR-CNN's 43.0 AP in standard 3× training schedule.