DeepSeek V4.1 Flash Object Detection Tested with Local RF-DETR Cross-Validation

On September 15, developer Maziyar Panahi shared an experiment: using DeepSeek V4.1 Flash to detect objects in 10 completely different photos, then cross-validating with a locally-run RF-DETR and keeping only the boxes both models agreed on—no box needed to be moved or manually fixed, and everything ran locally (the model was called via OpenRouter). He described it as too fun to stop.

Confirmed

• Total model call cost for all 10 photos was about $0.07
• Bird scenes (wings, waves, tiny targets) detected well
• Orange pile photo: only 21 of 28 visible candidate boxes passed RF-DETR consistency filtering, with some loose boxes appearing
• The airport was the hardest scene: only 5 of 41 visible candidate boxes passed filtering
• The author attached 4 best examples as a curated wrap-up

Why it matters

The experiment demonstrates a low-cost "VLM detection plus local specialized-model cross-validation" paradigm that yields usable boxes without human intervention. However, results on dense same-class objects (orange pile) and complex large scenes (airport) show that consistency filtering has clear limits under heavy occlusion and target density—real-world adoption still requires per-scene trade-offs.

2026-09-15 ~ 2026-09-15 · 5 related posts

• Zero Hand-Fixed Boxes: DeepSeek V4.1 Flash Does Local Object Detection Cross-Checked by RF-DETR — MaziyarPanahi · 2026-09-15
• [source] Object Detection with DeepSeek V4.1 Flash, Filtered by Local RF-DETR Agreement, Zero Hand-Fixes — MaziyarPanahi · 2026-09-15
• Orange Pile Stress Test: Only 21 of 28 Proposals Survive RF-DETR Agreement Filter — MaziyarPanahi · 2026-09-15
• 10-Photo Detection Run Cost Just $0.07; Airport Was the Toughest Scene at 5/41 — MaziyarPanahi · 2026-09-15
• Fully local object detection with DeepSeek V4.1 Flash, cross-checked by RF-DETR — MaziyarPanahi · 2026-09-15