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

On September 15, developer Maziyar Panahi shared a set of experiments: using DeepSeek V4.1 Flash to detect objects in 10 completely different photos, then cross-validating with only a locally run RF-DETR and keeping the boxes both models agreed on—without manually moving or fixing a single box, and with the entire pipeline running locally (the service called via OpenRouter). He called it "too fun to stop."

Confirmed

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

Why it matters

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

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

• [source] Zero Hand-Fixed Boxes: DeepSeek V4.1 Flash Does Local Object Detection Cross-Checked by RF-DETR — MaziyarPanahi · 2026-09-15
• Object Detection with DeepSeek V4.1 Flash, Filtered by Local RF-DETR Agreement, Zero Hand-Fixes — MaziyarPanahi · 2026-09-15
• [source] Orange Pile Stress Test: Only 21 of 28 Proposals Survive RF-DETR Agreement Filter — MaziyarPanahi · 2026-09-15
• [source] 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