Efficient Segment Anything
summary
EfficientSAM and SqueezeSAM are compact reworks of Segment Anything that keep quality while cutting compute: EfficientSAM uses mask image pretraining and a lightweight ViT to get 44.4 AP versus 46.5 AP for the original — about a 2-point drop — while using ~20× fewer parameters and running ~20× faster; a tiny EfficientSAM has 9.8M parameters versus 68M for FastSAM. SqueezeSAM is a fully convolutional model that is ~63× faster and ~32× smaller than SAM and improves whole-object segmentation. EfficientSAM-Video matches Sam2 (74.5% vs 74.7%) at >2× speed and runs ≈100–200 ms/frame on iPhone 15.
AI-written summary · about
chapters (8)
- 0:00Efficient segmentation models
- 1:55SAM architecture benefits
- 3:41Performance vs original SAM
- 7:38Squeeze SAM for segmentation
- 10:48Squeeze SAM efficiency
- 12:15Efficient SAM for video
- 13:41SAM Video performance
- 14:49Key takeaways
related talks
20mProductionizing On-device AI Models
22mEfficient Pretraining of Text-to-Image Models
15mTinySeg: Model Optimizing Framework for Image Segmentation on Tiny Embedded Systems
14mRegion-based Content Enhancement for Efficient Video Analytics at the Edge
01h 29mHuman-like Visual Search Application with Small Data
12mAccelerating On-Device ML Inference With ExecuTorch and Arm SME2
32mEverything you wanted to know about VLMs but were afraid to ask
32mTeaching Machines to See:Vision-Language Models Made Efficient with SmolVLM