DAM4SAMv2: An improved engineered version of DAM4SAM

Sebastian Janampa (Developer)

The University of New Mexico
Department of Electrical and Computer Engineering

---

DAM4SAMv2 is an improved version of DAM4SAM. We keep the great distractor-aware memory (DAM) consisting of two key parts: the recent appearances memory (RAM) and distractor-resolving memory (DRM).

DAM4SAM has not only surpassed SAM2 metrics, but it also addresses problems such as out-of-memory (OOM) errors caused by SAM2. Due to these factors, DAM4SAM has become a popular choice for creating segmentation/tracking ground-truth datasets. However, the official DAM4SAM implementation has two main drawbacks: 1) it only accepts bounding boxes as input, 2) single-object tracking, and 3) only uses the first frame of the video. Although the dam4sam-multi repo expands DAM4SAM capabilities for multi-object tracking, it does not address 1) and 2).

In this repo, we improve those drawbacks by developing DAM4SAMv2 (an improved engineered version of DAM4SAM). We allow users to select/edit/delete bounding boxes and points (which can also change the class of the points), use different frames from the video, and track multiple objects. For the objects, it is not required that they share the same frames, since not all objects appear in the first frame. This feature provides the user with more flexibility and facilitates the segmentation/tracking process.

demo_output.mp4

Video demos

We show some videos on how to use the repo in a local server and in Colab.

Local Server Demo

server_demo.mp4

Colab Notebook Demo

colab_demo.mp4

📝 TODO

• Create Python package.
• Report with evaluations (fps, memory usage, accuracy) on different datasets.

🚀 Updates

• [2025.10.04] DAM4SAMv2 code release.

Model Zoo

This repo supports the following models

• sam21pp-L
• sam21pp-B
• sam21pp-S
• sam21pp-T
• sam2pp-L
• sam2pp-B
• sam2pp-S
• sam2pp-T

Quick start

Set up

Method 1: git clone

git clone DAM4SAMv2
cd DAM4SAMv2
pip install -e .

Method 2: pip + git

pip install git+https://github.com/SebastianJanampa/DAM4SAMv2

Usage

Easy to run. For available models, see the Model Zoo section.

from dam4sam import DAM4SAM

tracker = DAM4SAM(model='sam21pp-L')
tracker.track(<video_file>)

track Function Parameters

This function runs the full tracking pipeline, handling everything from input validation to saving the final results.

• input_dir (str): Required. The path to the input video file or the directory containing your sequence of image frames.
• output_dir (str, optional): Specifies the directory where all output files (annotated frames, bounding boxes) will be saved. If None, the function runs in a visualization-only mode without saving files. Defaults to None.
• file_extension (str, optional): The file extension for the image frames (e.g., 'png', 'jpg') when input_dir is a directory. Defaults to 'jpg'.
• fp16 (bool, optional): If True, the model will use half-precision (FP16) for inference, which can improve performance on supported GPUs. Defaults to True.
• visualize (bool, optional): If True, a window will be displayed showing the live tracking process. Defaults to True.
• save_bboxes (bool, optional): If True, the bounding box coordinates for each tracked object are saved in .txt files, one for each frame. Defaults to True.
• save_frames (bool, optional): If True, the annotated frames with masks and bounding boxes will be saved as individual image files. Defaults to True.
• save_video (bool, optional): If True, a video will be generated from the annotated frames. This automatically enables save_frames if it's disabled. Defaults to True.
• out_video_name (str, optional): The desired filename for the output video. Defaults to None.
• exist_ok (bool, optional): If False, the function will create a new output directory with a numerical suffix (e.g., output2, output3) if the specified output_dir already exists. Defaults to False.

Platforms

This code was tested on a Linux machine with access to a GPU. We know that not everyone may have access to a GPU. That's why we also built this repo to support the Colab Environment. Nevertheless, if you use Colab, you will not be able to add points, but you will you will still retain all the other features. Our built-in COLAB GUI is also not very user-friendly, but we were unable to find an alternative solution since Colab does not support dynamic windows.

Citation

If you use this repo, please cite the DAM4SAM article. And if you like this repo, we would appreciate a 🌟

@InProceedings{dam4sam,
  author = {Videnovic, Jovana and Lukezic, Alan and Kristan, Matej},
  title = {A Distractor-Aware Memory for Visual Object Tracking with {SAM2}},
  booktitle = {Comp. Vis. Patt. Recognition},
  year = {2025}
}

Acknowledgement

This work is built upon SAM2, DAM4SAM, and dam4sam-multi. The GUIs were developed using GEMINI from GOOGLE. Big thanks, Google, for creating such a great multi-modal model.

✨ Feel free to contribute and reach out if you have any questions! ✨