SOLOv2: Dynamic and Fast Instance Segmentation

• DOI: 10.48550/arxiv.2003.10152
• Publication Date: 2020-01-01
• Authors: Xinlong Wang; Rufeng Zhang; Tao Kong; Lei Li; Chunhua Shen
• Keywords: Kernel (algebra); Feature (linguistics); Convolution (computer science); Matrix (chemical analysis); Code (set theory)

In this work, we aim at building a simple, direct, and fast instance segmentation framework with strong performance. We follow the principle of SOLO method Wang et al. "SOLO: segmenting objects by locations". Importantly, take one step further dynamically learning mask head object segmenter such that is conditioned on location. Specifically, branch decoupled into kernel feature branch, which are responsible for convolution convolved features respectively. Moreover, propose Matrix NMS (non maximum suppression) to significantly reduce inference time overhead due masks. Our performs parallel matrix operations in shot, yields better results. demonstrate simple direct system, outperforming few state-of-the-art methods both speed accuracy. A light-weight version SOLOv2 executes 31.3 FPS 37.1% AP. our results detection (from byproduct) panoptic show potential serve as new baseline many instance-level recognition tasks besides segmentation. Code available at: https://git.io/AdelaiDet