WildRoadBench - mAP: leaderboard

WildRoadBench (2026) asks vision-language models to box road damage in wild aerial images: water, cracks, potholes, debris and rare infrastructure defects. This board reports mAP averaged over IoU thresholds from 0.5 to 0.95, in percent, which rewards tight boxes; higher is better.

Metric: mAP (%) averaged over IoU thresholds 0.5 to 0.95, VLM Track, zero-shot box grounding on the full 1,061-image aerial set (1,394 ground-truth boxes); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 25 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro14.7#145
2Qwen 3.6 Plus12.4#238
3Gemini 3.1 Flash Lite11.4#371
4Qwen 2.5 VL 7B Instruct9#1902
5Claude Sonnet 4.68.7#173
6Qwen2.5-VL-32B-Instruct-AWQ8.6
7Qwen 3 VL 8B Instruct8.2#1029
8Qwen3-VL-32B-Instruct-AWQ7.9
9Qwen3.5-VL-27B-FP87.8
10Kimi K2.57#266
11GPT-5.4 (2026-03-05)6.9#92
12Qwen 3.5 VL 9B5.1
13Gemini 2.5 Flash5#463
14Qwen3-VL-32B-Thinking-AWQ4.9
15Qwen3.5-VL-35B-A3B-FP84.8

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=wildroadbench-map · How It Works · Data refreshed daily, snapshot 2026-11-21.