Research map / Detection and segmentation / CNN features
Improving Occlusion and Hard Negative Handling for Single-Stage Pedestrian Detectors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add soft part-score occlusion handling (ablation) | SqueezeDet+, YOLOv2, SSD, and DSSD · Caltech test, Height ≥50, All | log-average miss rate: 32.83 -> 30.18; 29.35 -> 27.16; 25.18 -> 22.53; 20.53 -> 18.58 | Helped |
| Add grid classifiers with adjusted confidence (ablation) | SqueezeDet+ and YOLOv2 · Caltech test, Height ≥50, All | log-average miss rate: 32.83 -> 28.72; 29.35 -> 27.65 | Helped |
| Add grid classifiers for training only (ablation) | SSD and DSSD · Caltech test, Height ≥50, All | log-average miss rate: 25.18 -> 23.79; 20.53 -> 18.20 | Helped |
| Add soft part-score occlusion handling (ablation) | SqueezeDet+, YOLOv2, SSD, and DSSD · CityPersons val, Height ≥50, All | log-average miss rate: 43.90 -> 41.90; 38.01 -> 36.36; 35.61 -> 33.95; 34.37 -> 33.16 | Helped |
| Add grid classifiers with adjusted confidence (ablation) | SqueezeDet+ and YOLOv2 · CityPersons val, Height ≥50, All | log-average miss rate: 43.90 -> 41.92; 38.01 -> 36.76 | Helped |
| Add grid classifiers for training only (ablation) | SSD and DSSD · CityPersons val, Height ≥50, All | log-average miss rate: 35.61 -> 34.32; 34.37 -> 31.67 | Helped |
| Jointly add part-score occlusion handling and grid classifiers (ablation) | SqueezeDet+, YOLOv2, SSD, and DSSD · Caltech test, Height ≥50, All | log-average miss rate: 32.83 -> 28.29; 29.35 -> 26.61; 25.18 -> 23.54; 20.53 -> 19.38 | Helped |
| Jointly add part-score occlusion handling and grid classifiers (ablation) | SqueezeDet+, YOLOv2, SSD, and DSSD · CityPersons val, Height ≥50, All | log-average miss rate: 43.90 -> 40.88; 38.01 -> 34.09; 35.61 -> 33.52; 34.37 -> 31.71 | Helped |
| Use max part score instead of the base confidence (ablation) | SqueezeDet+, YOLOv2, SSD, and DSSD · Caltech test, Height ≥50, Heavy occlusion | log-average miss rate: 63.65 -> 56.60; 57.55 -> 53.90; 53.80 -> 50.02; 44.13 -> 44.81 | Helped |
| Use soft part score instead of the base confidence (ablation) | SqueezeDet+, YOLOv2, SSD, and DSSD · Caltech test, Height ≥50, Heavy occlusion | log-average miss rate: 63.65 -> 59.87; 57.55 -> 57.02; 53.80 -> 50.46; 44.13 -> 44.11 | Mixed |
| Use adjusted grid-classifier confidence at inference (ablation) | SSD and DSSD · Caltech test, Height ≥50, All | log-average miss rate: 25.18 -> 23.79 | Hurt |
| Jointly add part-score handling and grid classifiers (ablation) | SSD · Caltech test, Height ≥50, Heavy occlusion | log-average miss rate: 53.80 -> 51.53 | Hurt |
Closest papers by meaning
- How Far Are We From Solving Pedestrian Detection? (CVPR 2016 · similarity 0.88)
- PedHunter: Occlusion Robust Pedestrian Detector in Crowded Scenes (AAAI 2020 · similarity 0.88)
- Occluded Pedestrian Detection Through Guided Attention in CNNs (CVPR 2018 · similarity 0.87)
- Temporal-Context Enhanced Detection of Heavily Occluded Pedestrians (CVPR 2020 · similarity 0.86)
- Generalizable Pedestrian Detection: The Elephant in the Room (CVPR 2021 · similarity 0.86)
- Multi-View Pedestrian Occupancy Prediction with a Novel Synthetic Dataset (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.