Vision AI in 2025
summary
Argues that current vision models “aren’t smart” and shows concrete fixes: swapping a Dinov2 backbone into a transformer detector yields +5 mAP vs the convolutional baseline; introduces the RFDTOR model for real-time edge detection and the RF100VL dataset (100 curated datasets, 10-shot canonical splits, many imaging domains) which reveals huge domain-adaptability gains from strong pretraining and exposes that state-of-the-art VLMs score worse than tiny YOLOv8 when fine-tuned on 10 examples per class.
AI-written summary · about
chapters (8)
- 0:00Introduction to Vision AI
- 1:45Vision evals are saturated
- 3:11Vision models aren't smart
- 6:40Aligning vision and language features
- 8:24Transformers leverage big pre-training
- 9:05Introducing RF DETR model
- 10:31Introducing RF100VL benchmark
- 13:22Summary and Q&A
related talks
17mHow Transformers Finally Ate Vision
33mSeeing Through the Lens of Transformers: A New Era in Computer Vision
01h 08mMastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques
41mBridging Pixels and Words: The Transformer Vision Revolution
41mDeep Learning to Solve Challenging Problems
29mApplied Custom Vision
14mA Primer to Modern Day Computer Vision
42mState of the Art of Innovation in Computer Vision