Rethinking Space-Time Networks with Improved Memory Coverage for Efficient Video Object Segmentation

NeurIPS 2021  ·  Ho Kei Cheng, Yu-Wing Tai, Chi-Keung Tang ·

This paper presents a simple yet effective approach to modeling space-time correspondences in the context of video object segmentation. Unlike most existing approaches, we establish correspondences directly between frames without re-encoding the mask features for every object, leading to a highly efficient and robust framework. With the correspondences, every node in the current query frame is inferred by aggregating features from the past in an associative fashion. We cast the aggregation process as a voting problem and find that the existing inner-product affinity leads to poor use of memory with a small (fixed) subset of memory nodes dominating the votes, regardless of the query. In light of this phenomenon, we propose using the negative squared Euclidean distance instead to compute the affinities. We validated that every memory node now has a chance to contribute, and experimentally showed that such diversified voting is beneficial to both memory efficiency and inference accuracy. The synergy of correspondence networks and diversified voting works exceedingly well, achieves new state-of-the-art results on both DAVIS and YouTubeVOS datasets while running significantly faster at 20+ FPS for multiple objects without bells and whistles.

PDF Abstract NeurIPS 2021 PDF NeurIPS 2021 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Semi-Supervised Video Object Segmentation DAVIS 2016 STCN Jaccard (Mean) 90.4 # 17
Jaccard (Recall) 98.1 # 1
Jaccard (Decay) 4.1 # 31
F-measure (Mean) 93.0 # 18
F-measure (Recall) 97.1 # 1
F-measure (Decay) 4.3 # 30
J&F 91.7 # 17
Speed (FPS) 26.9 # 20
Semi-Supervised Video Object Segmentation DAVIS 2017 (test-dev) STCN J&F 79.9 # 18
Jaccard (Mean) 76.3 # 18
Jaccard (Recall) 85.5 # 1
Jaccard (Decay) 10.5 # 1
F-measure (Mean) 83.5 # 19
F-measure (Recall) 89.7 # 1
F-measure (Decay) 10.3 # 1
Semi-Supervised Video Object Segmentation DAVIS 2017 (val) STCN Jaccard (Mean) 82.0 # 23
Jaccard (Recall) 91.3 # 2
Jaccard (Decay) 6.2 # 1
F-measure (Mean) 88.6 # 17
F-measure (Recall) 94.6 # 1
F-measure (Decay) 85.3 # 23
J&F 85.3 # 20
Speed (FPS) 20.2 # 20
Semi-Supervised Video Object Segmentation MOSE STCN J&F 50.8 # 15
J 46.6 # 15
F 55.0 # 14
Semi-Supervised Video Object Segmentation YouTube-VOS 2018 STCN F-Measure (Seen) 87.9 # 23
F-Measure (Unseen) 87.3 # 13
Jaccard (Seen) 83.2 # 23
Jaccard (Unseen) 79.0 # 13
Semi-Supervised Video Object Segmentation YouTube-VOS 2019 STCN Overall 84.2 # 16
Jaccard (Seen) 82.6 # 17
Jaccard (Unseen) 79.4 # 12
F-Measure (Seen) 87.0 # 17
F-Measure (Unseen) 87.7 # 11
Semi-Supervised Video Object Segmentation YouTube-VOS 2019 STCN (MS) Overall 85.2 # 9
Jaccard (Seen) 83.5 # 13
Jaccard (Unseen) 80.8 # 7
F-Measure (Seen) 87.8 # 15
F-Measure (Unseen) 88.8 # 7
Video Object Segmentation YouTube-VOS 2019 STCN Mean Jaccard & F-Measure 82.7 # 7
Jaccard (Seen) 81.1 # 8
Jaccard (Unseen) 78.2 # 6
F-Measure (Seen) 85.4 # 8
F-Measure (Unseen) 85.9 # 6

Methods


No methods listed for this paper. Add relevant methods here