Dataset Viewer
Auto-converted to Parquet Duplicate
passed
bool
validated_images
int64
validated_binary_masks
int64
validated_instance_masks
int64
source_images_readable_checked
bool
original_dimensions_checked
bool
binary_png_values_checked
bool
rle_area_bbox_and_union_checked
bool
coco_aggregation_checked
bool
seconds
float64
note
string
true
3,155
6,310
7,301
true
true
true
true
true
54.82
Artifact consistency validation; it does not measure segmentation accuracy against manual labels.

SAM 3 person and seat belt segmentation

This is the latest folder-input rerun on 3,155 vehicle-cabin images. The SAM 3 source revision was f92a5efc4001ebfe9d050d43e61f6fb10e02d73c; the checkpoint SHA-256 is 9999e2341ceef5e136daa386eecb55cb414446a00ac2b55eb2dfd2f7c3cf8c9e.

  • segment_archive.py: SAM 3 inference from an image folder. Its historical filename does not imply ZIP support.
  • visualize_person_belt_split.py: geometry-only shoulder-entry approximation using the largest person instance and smoothed row boundaries; no additional model is loaded.
  • validate_results.py, visualize_all.py: artifact validation and local gallery generation.
  • sam3/overlays/: 3,155 JPEG overlays (green person, orange seat belt).
  • sam3/masks/person/, sam3/masks/seat_belt/: 3,155 binary PNG masks per class (background 0, foreground 255).
  • sam3/annotations/, sam3/annotations_coco.json: per-image and aggregated COCO instance segmentations.
  • sam3/summary.csv, sam3/summary.json, sam3/validation.json: image-level counts, source provenance and artifact consistency checks.
  • shoulder_split/overlays/: 3,155 JPEG visualizations (yellow person box, orange belt inside the estimated body silhouette, blue belt outside).
  • shoulder_split/masks/inside/, shoulder_split/masks/outside/: 3,155 binary PNG masks per split; they are disjoint and their union equals the original belt mask.
  • shoulder_split/summary.csv, shoulder_split/preview.jpg: pixel counts and a 16-image overview.

The original unmodified images and SAM 3 checkpoint are not included. The overlays still show the original scenes. These are automatic predictions, not manual labels or a seat-belt-use compliance decision. The shoulder boundary is an approximation, not a pose keypoint.

python segment_archive.py --input /path/to/images --output /path/to/results
python segment_archive.py --input /path/to/images --output /path/to/results --summarize
python validate_results.py --output /path/to/results
python visualize_person_belt_split.py --results /path/to/results
Downloads last month
126