sosa123454321/cekicin-damage-benchmark

Çekicin car damage benchmark

Which model should the Çekicin roadside-assistance app use to read damage from a customer's photo? This dataset compares the two small models that run on the phone / in the browser with two big vision-language models that run in the cloud, on the same 100 labelled photos.

• 🧪 Try the models: Çekicin AI Lab
• 🔌 API used by the app: https://cekicin-ai-api.elasa2next.workers.dev (OpenAPI)
• 📚 Cost rules: cekicin-ai-knowledge

Round 4 (2026-10-17): photo gallery + 12 real scenarios

• test1000/gallery/: 48 random test1000 photos with the boxes of 8 models → model-test page.
• scenarios12/: 12 Wikimedia Commons photos in different situations (close-ups, severe crash, night, snow, hail, intact car), on-device vs cloud, per-step counts and max score per type → scenarios page + explainer video. Qualitative, not a benchmark. On-device: 4 correct, 1 partly, 1 wrong, 1 false alarm, 5 missed (all wide/distant scenes); cloud: 10 correct, 2 partly (no tow advised in two severe crashes).

Round 3 (2026-10-16): test1000 — debugging + 5 Hugging Face models

Folder test1000/: 1000 photos (888 damaged from AutoDamageIQ val / VehiDE + 112 intact cars from Stanford Cars), every detection of every model, scripts and the report (benchmark-1000 page, mirror).

model	macro-F1	damage recall	false alarm on intact cars	AUROC	CPU ms
Çekicin CarDD YOLOv8s-seg int8 (on-device)	0.52	82 %	41 % → 28 % with the v1.2.2 fix	0.78	243
shanexf/car-damage-yolo11m-ft	0.51	76 %	14 %	0.89	453
beingamit99/car_damage_detection (ViT)	0.46	98 %	95 %	0.82	169
Çekicin parts YOLO11n (on-device)	0.25	70 %	46 %	0.67	45
tatar01230/car-damage-yolo11n	0.18	78 %	52 %	0.72	34
Naiscorp/car-damage-detection (300-photo subset)	0.64	83 %	12 %	0.89	4364

Debug result: the browser pipeline equals the original PyTorch weights; the problem was training data (damaged cars only → false alarms on intact headlights/tyres), fixed with per-class floors (lamp/tyre 0.6). See test1000/README.md.

Round 2: hard cases and the tow decision

hard/ adds 13 clean cars, 40 broken lamps, 14 flat tyres, 8 severe crashes and night/blur variants, all scored on the tow decision.

With the v1.2 hybrid (CarDD + Llama 4, cloud veto), tow recall is 21/22 and false tows are 0/16 (v1.1 logic: 20/22 and 3/16). Night photos halve accuracy and blur destroys it, so the app now checks photo quality first.

Results (image-level F1 per damage type)

Model	Where it runs	Dent	Scratch	Crack	Glass	Macro F1 (95% CI)	Any damage found	Median time
YOLO11n parts	phone/browser	0.55	–	–	0.48	0.26	66%	0.1 s (CPU)
CarDD YOLOv8s-seg, 8-bit	phone/browser	0.45	0.64	0.47	0.65	0.56 (0.47–0.62)	82%	0.4 s (CPU)
Llama 4 Scout 17B-16E	cloud (Workers AI)	0.59	0.69	0.37	0.70	0.59 (0.52–0.65)	96%	3.1 s
Mistral Small 3.1 24B	cloud (Workers AI)	0.57	0.60	0.27	0.83	0.57 (0.50–0.63)	91%	4.1 s
CarDD ∪ Llama 4 (hybrid)	phone + cloud	0.59	0.76	0.47	0.79	0.65 (0.58–0.71)	100%	~3.5 s
Vote 2-of-3 (CarDD, Llama, Mistral)	phone + 2× cloud	0.53	0.70	0.35	0.77	0.59	83%	—

Decision: the app uses the hybrid — CarDD on the device (instant, offline, also draws the damage outline) and Llama 4 Scout in the cloud, merged by the API. The hybrid beats Llama-only in 100% of 2,000 paired bootstrap resamples; the single models are statistically tied. Cost: ~32 Workers AI neurons per photo ≈ US$0.00035 per cloud assessment on the paid plan, free for ~300 photos/day on the free plan.

Main weaknesses found: cracks are hard for every model (best F1 0.47); the parts model cannot see scratches or cracks at all; LLMs under-report dents (recall 0.50) and sometimes call a cracked bumper a scratch.

Data

• samples.csv — 100 images from the validation split of tugberkkalay/autodamageiq-vehicle-damage-dataset (CC-BY-4.0; images originate from VehiDE). Stratified: ≥30 glass, ≥30 crack, then random. Label counts (images): scratch 58, dent 44, glass 37, crack 36. Images are referenced, not copied (download_images.py). None come from CarDD, so the CarDD model is tested out-of-domain (no leakage).
• predictions.jsonl — every model output (boxes + scores for ONNX models; parsed JSON + raw text, latency and neurons for cloud models).
• scores.csv — precision, recall, F1 per class, macro F1, "any damage" recall, exact-match rate, latency.
• prompt.txt — the exact production prompt (same as the API).
• Scripts: download_images.py, run_local_models.py, run_cloud_models.py (needs CLOUDFLARE_API_TOKEN, CLOUDFLARE_ACCOUNT_ID), score.py, bootstrap.py.

Method

• Task = image-level presence of each damage type (dent, scratch, crack, glass_shatter). Lamp and tyre are absent from this split and not scored.
• Parts-model classes mapped to damage types (*-dent→dent, windscreen→glass, lamp→lamp_broken).
• ONNX models: onnxruntime CPU, confidence 0.25, 640 px. Cloud models: temperature 0, image resized to 768 px.
• Hybrid = union of CarDD and Llama 4 findings. Vote = present if ≥2 of 3 models agree.

Limitations

Round 1 (100 images): 100 images is small (see confidence intervals). Ground truth comes from a merged dataset with some noisy scratch/dent labels. No undamaged cars were included in round 1; round 3 (test1000) adds 112 intact cars to measure false alarms. Not a substitute for an expert. Prices in the API are illustrative.

Licences

Benchmark files: CC-BY-4.0 (attribution: AutoDamageIQ / VehiDE authors and Çekicin). CarDD model weights: AGPL-3.0, CarDD data non-commercial research only. Llama 4: Llama 4 Community License. Mistral Small 3.1: Apache-2.0.