Pyramid Dilated Deeper ConvLSTM for Video Salient Object Detection
This paper proposes a fast video salient object detection model, based on a novel recurrent network architecture, named Pyramid Dilated Bidirectional ConvLSTM (PDB-ConvLSTM). A Pyramid Dilated Convolution (PDC) module is first designed for simultaneously extracting spatial features at multiple scales. These spatial features are then concatenated and fed into an extended Deeper Bidirectional ConvLSTM (DB-ConvLSTM) to learn spatiotemporal information. Forward and backward ConvLSTM units are placed in two layers and connected in a cascaded way, encouraging information flow between the bi-directional streams and leading to deeper feature extraction. We further augment DB-ConvLSTM with a PDC-like structure, by adopting several dilated DB-ConvLSTMs to extract multi-scale spatiotemporal information. Extensive experimental results show that our method outperforms previous video saliency models in a large margin, with a real-time speed of 20 fps on a single GPU. With unsupervised video object segmentation as an example application, the proposed model (with a CRF-based post-process) achieves state-of-the-art results on two popular benchmarks, well demonstrating its superior performance and high applicability.
PDF AbstractCode
Results from the Paper
Ranked #1 on Video Salient Object Detection on UVSD (using extra training data)
Task | Dataset | Model | Metric Name | Metric Value | Global Rank | Uses Extra Training Data |
Benchmark |
---|---|---|---|---|---|---|---|
Video Salient Object Detection | DAVIS-2016 | PDB | S-Measure | 0.882 | # 5 | ||
MAX E-MEASURE | 0.951 | # 2 | |||||
AVERAGE MAE | 0.028 | # 7 | |||||
Unsupervised Video Object Segmentation | DAVIS 2016 val | PDB | G | 75.9 | # 23 | ||
J | 77.2 | # 23 | |||||
F | 74.5 | # 22 | |||||
Unsupervised Video Object Segmentation | DAVIS 2017 (test-dev) | PDB | J&F | 40.4 | # 5 | ||
Jaccard (Mean) | 37.7 | # 4 | |||||
Jaccard (Recall) | 42.6 | # 4 | |||||
Jaccard (Decay) | 4.0 | # 4 | |||||
F-measure (Mean) | 43.0 | # 4 | |||||
F-measure (Recall) | 44.6 | # 4 | |||||
F-measure (Decay) | 3.7 | # 3 | |||||
Unsupervised Video Object Segmentation | DAVIS 2017 (val) | PDB | J&F | 55.1 | # 9 | ||
Jaccard (Mean) | 53.2 | # 9 | |||||
Jaccard (Recall) | 58.9 | # 7 | |||||
F-measure (Mean) | 57.0 | # 9 | |||||
F-measure (Recall) | 60.2 | # 7 | |||||
Video Salient Object Detection | DAVSOD-Difficult20 | PDB | S-Measure | 0.608 | # 2 | ||
max E-measure | 0.678 | # 4 | |||||
Average MAE | 0.107 | # 1 | |||||
Video Salient Object Detection | DAVSOD-easy35 | PDB | S-Measure | 0.706 | # 2 | ||
max F-Measure | 0.591 | # 2 | |||||
max E-Measure | 0.749 | # 3 | |||||
Average MAE | 0.114 | # 4 | |||||
Video Salient Object Detection | DAVSOD-Normal25 | PDB | S-Measure | 0.649 | # 2 | ||
max E-measure | 0.698 | # 3 | |||||
Average MAE | 0.132 | # 4 | |||||
Video Salient Object Detection | FBMS-59 | PDB | S-Measure | 0.851 | # 5 | ||
AVERAGE MAE | 0.064 | # 5 | |||||
MAX F-MEASURE | 0.821 | # 4 | |||||
Unsupervised Video Object Segmentation | FBMS test | PDB | J | 74.0 | # 11 | ||
Video Salient Object Detection | MCL | PDB | S-Measure | 0.856 | # 1 | ||
MAX E-MEASURE | 0.911 | # 1 | |||||
AVERAGE MAE | 0.021 | # 8 | |||||
Video Salient Object Detection | SegTrack v2 | PDB | S-Measure | 0.864 | # 2 | ||
AVERAGE MAE | 0.024 | # 3 | |||||
max E-measure | 0.935 | # 1 | |||||
Video Salient Object Detection | UVSD | PDB | S-Measure | 0.901 | # 1 | ||
max E-measure | 0.975 | # 1 | |||||
Average MAE | 0.018 | # 1 | |||||
Video Salient Object Detection | ViSal | PDB | S-Measure | 0.907 | # 3 | ||
max E-measure | 0.846 | # 6 | |||||
Average MAE | 0.032 | # 3 | |||||
Video Salient Object Detection | VOS-T | PDB | S-Measure | 0.818 | # 3 | ||
max E-measure | 0.837 | # 3 | |||||
Average MAE | 0.078 | # 3 | |||||
Unsupervised Video Object Segmentation | YouTube-Objects | PDB | J | 65.5 | # 11 |