Description
The Caltech Pedestrian Dataset (Dollar et al.) is a large pedestrian-detection benchmark consisting of approximately 10 hours of 640x480 video recorded from a vehicle-mounted monocular camera in regular traffic. It has bounding-box annotations and an established evaluation toolkit, and remains the canonical urban pedestrian-detection benchmark.
Modality / size
- Modality: 640x480 RGB video from a vehicle-mounted camera.
- Subjects / scenarios: ~10 hours of urban driving footage.
- Labels: pedestrian bounding boxes with occlusion / visibility annotations.
Used by (papers)
- Used by detector pre-training studies in the vault, often paired with
INRIA pedestrian. - DeepCascade and similar detectors are trained on Caltech + auxiliary datasets.
Notes
- Public dataset; standard release from the Caltech Vision lab.