Vision AI in 2025 — Peter Robicheaux, Roboflow

Summary

• Roboflow's ML lead argues VLMs still can't see — Claude 3.5/4 guesses watch times and gets school-bus orientation wrong on MMVP, hallucinating supporting details.
• Root cause: CLIP-style caption contrastive pre-training can't distinguish image pairs whose visual differences (dog pose) aren't captioned; DinoV2 self-supervised pre-training does discriminate.
• Object-detection field hasn't adopted big pre-training: YOLOv8 (CNN) gains <0.2 mAP from Objects365 pre-training, while LW-DETR (transformer) gains 5-7 mAP.
• Roboflow announces RF-DETR, a real-time detector swapping in a DinoV2 backbone — close to SOTA on COCO and far ahead on the new RF100-VL benchmark for domain adaptation.
• Introduces RF100-VL: 100 curated detection datasets covering aerial views and unusual modalities to measure visual intelligence beyond saturated COCO.