基于动态采样对偶可变形网络的实时视频实例分割

宋一然,周千寓,邵志文,易冉,马利庄

Dynamic sampling dual deformable network for online video instance segmentation

Yiran SONG,Qianyu ZHOU,Zhiwen SHAO,Ran YI,Lizhuang MA

表 1 基于YouTube-VIS 2019验证集的视频实例分割方法的比较

Tab.1 Comparisons of video instance segmentation on YouTube-VIS 2019 validation dataset

方法 mAP/% AP50/% AP75/% MaskTrack R-CNN 50[8] 30.3 51.1 32.6 MaskTrack R-CNN 101[8] 41.8 53.0 33.6 MaskProp 50 [10] 40.0 — 42.9 MaskProp 101 [10] 42.5 — 45.6 *VisTR 50 [11] 36.2 59.8 36.9 *VisTR 101 [11] 40.1 64.0 45.0 CrossVIS 50 [3] 36.3 56.8 38.9 CrossVIS 101[3] 36.6 57.3 39.7 CompFeat 50 [31] 35.3 56.0 38.6 *IFC 50 [22] 41.0 62.1 45.4 STC [32] 36.7 57.2 38.6 VSTAM [33] 39.0 62.9 41.8 SipMask 50 [2] 33.7 54.1 35.8 DSDDN 50 37.5 59.1 41.9 DSDDN 101 39.1 60.7 43.5