# Efficient Track Anything

**Authors:** Yunyang Xiong, Chong Zhou, Xiaoyu Xiang, Lemeng Wu, Chenchen Zhu, Zechun Liu, Saksham Suri, Balakrishnan Varadarajan, Ramya Akula, Forrest Iandola, Raghuraman Krishnamoorthi, Bilge Soran, Vikas Chandra  
**arXiv:** [2411.18933](https://arxiv.org/abs/2411.18933)  
**Date:** 2024-11-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2411.18933v1/

## Key points

- Efficient Track Anything research paper by Yunyang Xiong et al.
- Limitations of Segment Anything Model 2 (SAM 2) in real-world applications due to high computation complexity
- Proposal of EfficientTAMs - lightweight track anything models for high-quality results with low latency and model size
- Utilization of plain Vision Transformer (ViT) as an image encoder and efficient memory module for frame feature extraction and memory computation
- Training on SA-1B and SA-V datasets for video object segmentation and track anything tasks
- Evaluation on multiple video segmentation benchmarks showing promising results
- EfficientTAM with vanilla ViT performs comparably to SAM 2 model with speedup and parameter reduction on A100 GPUs
- Outperformance of original SAM in segment anything image tasks with significant speedup and parameter reduction on A100 GPUs
- Efficiency demonstrated even on mobile devices like iPhone 15 Pro Max, running at approximately 10 frames per second for video object segmentation
- Significance of developing lightweight models like EfficientTAMs for efficient video object segmentation tasks

## Summary

Efficient Track Anything is a research paper authored by Yunyang Xiong, Chong Zhou, Xiaoyu Xiang, Lemeng Wu, Chenchen Zhu, Zechun Liu, Saksham Suri, Balakrishnan Varadarajan, Ramya Akula, Forrest Iandola,
Raghuraman Krishnamoorthi,Bilge Soran,and Vikas Chandra. The paper discusses the limitations of Segment Anything Model 2 (SAM 2) in real-world applications due to the high computation complexity of its multistage image encoder and memory module. To address this issue,the authors propose EfficientTAMs - lightweight track anything models that offer high-quality results with low latency and model size. EfficientTAMs are built on the concept of revisiting the plain Vision Transformer (ViT) as an image encoder for video object segmentation and introducing an efficient memory module to reduce complexity in frame feature extraction and memory computation for current frame segmentation. The authors utilize vanilla lightweight ViTs and the efficient memory module to train the models on SA-1B and SA-V datasets for video object segmentation and track anything tasks. The evaluation of EfficientTAMs on multiple video segmentation benchmarks including semi-supervised VOS and promptable video segmentation shows promising results. The proposed EfficientTAM with vanilla ViT performs comparably to SAM 2 model (HieraB+SAM 2) with approximately a 2x speedup on A100 GPUs and a 2.4x parameter reduction.In segment anything image tasks,EfficientTAMs outperform original SAM with around a 20x speedup on A100 GPUs and a 20x parameter reduction.Notably,EfficientTAMs demonstrate their efficiency even on mobile devices such as iPhone 15 Pro Max by running at approximately 10 frames per second for performing video object segmentation with reasonable quality. This highlights the capability of small models like EfficientTAMs for on-device video object segmentation applications. Overall,the research presented in Efficient Track Anything showcases a significant advancement in developing lightweight models for efficient video object segmentation tasks.