thirdeyelabs/indian-road-benchmark

Indian Road Detection Benchmark (preview)

904 dashcam frames from Delhi NCR, each labelled twice, independently, by our in-house annotation team. Off-the-shelf detectors keep about 59% of their COCO accuracy on these frames, and the best of 20 misses half the auto-rickshaws. The full leaderboard is at thirdeyelabs.ai/leaderboard.

What is in it

Split	Frames	Day	Dusk or dawn	Night	Rain	Labels
dev	450	160	60	230	54	public, in annotations/dev.json
test	454	160	59	235	56	held back; we score your predictions

• images/dev/ and images/test/: the frames, 1920 x 994 JPEG.
• annotations/dev.json: COCO format, 2,705 agreed boxes and 2,730 ignore regions.
• annotations/test_images.json: the test frames, with ids, but no boxes.
• evaluate.py: the scorer behind the leaderboard. It needs only numpy.

We split by drive, so no clip has frames on both sides, and we balanced conditions and class counts between the halves.

Labels

Our in-house team labelled every frame twice, in two independent passes on our own annotation platform. A box both passes drew is ground truth (iscrowd: 0). A box only one pass drew is an ignore region (iscrowd: 1): a model that finds it gains nothing, and a model that misses it loses nothing.

Classes: car, truck, bus, autorickshaw, motorcycle, bicycle, rider, person, animal, traffic light, traffic sign, vehicle fallback. Riders have their own box, separate from the motorcycle or bicycle they ride. Boxes are COCO [x, y, width, height] in pixels.

Every image also carries time_of_day (day, dusk, night) and rain (true or false), both checked by a person, plus frame_id and clip_id.

Scoring

python evaluate.py annotations/dev.json predictions.json

predictions.json is a list in the COCO results format:

[{"image_id": 12, "category_name": "car", "bbox": [x, y, width, height], "score": 0.91}, ...]

Use COCO class names, or our own names. A numeric category_id is read as a COCO id. Keep up to 100 boxes per image, low-confidence ones included.

The score is COCO-style mAP from IoU 0.50 to 0.95 over the eight classes a COCO model can name: person, bicycle, car, motorcycle, bus, truck, traffic light and animal. A person box counts for riders too. Near counts only objects at least 32 px tall. Each half is small, so differences under about 2 points on one half are noise. The leaderboard uses all 904 frames.

Baselines on dev

Published COCO weights, default input size, up to 100 boxes per frame above 1% confidence, nothing fine-tuned.

Model	Score	AP50	Near	Day	Dusk	Night	Rain
YOLO26x	34.2	57.5	35.4	40.0	31.8	31.2	41.7
YOLO11x	32.2	52.5	33.5	38.2	30.9	28.8	40.8
YOLOv8x	32.1	52.6	33.3	38.5	28.9	29.0	33.5
YOLOv9e	32.0	54.5	33.2	39.3	32.5	27.9	37.7
YOLO12x	32.0	53.2	33.4	36.9	31.6	29.7	37.8
YOLO26m	31.8	51.0	33.0	39.1	29.9	27.5	31.4
RT-DETR-X	31.3	53.7	32.4	36.6	30.5	29.3	38.1
YOLOv5x	31.0	51.5	32.3	37.2	30.1	27.3	36.0
YOLOv9c	31.0	50.0	32.4	35.7	28.5	28.2	27.8
YOLOv10x	30.8	50.1	32.1	36.6	27.7	27.7	35.6
RT-DETR-L	30.7	53.1	31.9	35.8	27.6	29.4	31.0
YOLO11m	30.7	50.5	32.1	36.5	28.9	27.6	35.6
YOLOv10m	30.5	50.1	31.7	37.3	28.1	27.4	36.4
YOLOv8m	30.1	49.6	31.2	34.5	27.6	28.0	34.8
Faster R-CNN v2	29.9	54.7	31.2	35.4	29.1	26.5	26.9
YOLO12m	29.3	48.7	30.5	32.9	31.7	26.9	36.2
RetinaNet v2	27.3	49.3	28.5	35.3	26.0	22.9	21.9
YOLOv5s	24.1	40.0	25.2	28.0	20.4	22.9	23.3
YOLO11n	19.7	34.3	20.6	23.0	14.7	20.1	27.0
YOLOv8n	19.3	33.0	20.2	21.6	15.9	19.1	16.8

Get scored on the test half

Run your model on images/test/, save predictions in the format above using the ids in annotations/test_images.json, and send the file to shivam@thirdeyelabs.ai. We reply with your scores and, if you like, add your model to the leaderboard. We also run private evaluations on the roads and conditions you ship into.

Privacy

We blur faces, number plates and numbers painted on vehicles. If you spot one we missed, write to us and we will fix it.

Source and licence

The frames come from our open dataset, thirdeyelabs/indian-road-dataset. Both are released under CC BY 4.0.

Preview status

Nobody has yet settled the boxes only one pass drew. Once someone adjudicates them, scores may move by a point or two.

Citation

@misc{thirdeye2026indianroad,
  title  = {Indian Road Detection Benchmark},
  author = {Third Eye Labs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/thirdeyelabs/indian-road-benchmark}
}