FindIt - Video Detection: leaderboard
Video detection: the model localizes the queried object in every one of 2 or 8 frames sampled from RoboTools and iGround clips. FindIt (2026), a format-informed localization benchmark for multimodal LLMs, scores each model with the bounding-box and output format it handles best, as F1 at IoU 0.5 averaged over the datasets, in percent; higher is better.
Metric: Average F1@0.5 (%) over video detection (localize the target in every sampled frame) on RoboTools instance detection and iGround object detection clips at 2 and 8 uniformly sampled frames; each model scored with the box representation, text or JSON output and JSON key that a two-stage probe on Pascal VOC found best for it; boxes Hungarian-matched to the ground truth, a match counts at IoU 0.5 or more (with the right label for multi-label queries); unparseable outputs count as empty; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Non-reasoning) | 31.4 | #123 (GPT-5.4) |
| 2 | Qwen 3.5 9B (Thinking) | 30.2 | #803 (Qwen 3.5 9B) |
| 3 | Gemini 2.5 Flash (Non-reasoning) | 28.9 | #453 (Gemini 2.5 Flash) |
| 4 | Qwen 3 VL 8B Instruct | 25.8 | #981 |
| 5 | InternVL3-8B | 17.9 | #1688 |
| 6 | Gemma 4 E4B | 17.2 | #2214 |
| 7 | Claude Sonnet 4.5 | 11.8 | #250 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=findit-video-detection · How It Works · Data refreshed daily, snapshot 2026-11-13.