Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, Joon-Young Lee

University of Illinois Urbana-Champaign and Adobe

ICCV 2023

[arXiV] [PDF] [Project Page]

1. Provide long-term, open-vocabulary video segmentation with text-prompts out-of-the-box.
2. Fairly easy to integrate your own image model! Wouldn't you or your reviewers be interested in seeing examples where your image model also works well on videos 😏? No finetuning is needed!

Note (Sep 12 2023): We have improved automatic video segmentation by not querying the points in segmented regions. We correspondingly increased the number of query points per side to 64 and deprecated the "engulf" mode. The old code can be found in the "legacy_engulf" branch. The new code should run a lot faster and capture smaller objects. The text-prompted mode is still recommended for better results.

Note (Sep 11 2023): We have removed the "pluralize" option as it works weirdly sometimes with GroundingDINO. If needed, please pluralize the prompt yourself.

We develop a decoupled video segmentation approach (DEVA), composed of task-specific image-level segmentation and class/task-agnostic bi-directional temporal propagation. Due to this design, we only need an image-level model for the target task and a universal temporal propagation model which is trained once and generalizes across tasks. To effectively combine these two modules, we propose a (semi-)online fusion of segmentation hypotheses from different frames to generate a coherent segmentation. We show that this decoupled formulation compares favorably to end-to-end approaches in several tasks, most notably in large-vocabulary video panoptic segmentation and open-world video segmentation.

Source: https://www.youtube.com/watch?v=FM9SemMfknA

Source: https://youtu.be/FbK3SL97zf8

Source: https://youtu.be/couz1CrlTdQ

Source: DAVIS 2017 validation set "soapbox"

Source: https://youtu.be/FQQaSyH9hZI

Tested on Ubuntu only. For installation on Windows WSL2, refer to hkchengrex#20 (thanks @21pl).

Prerequisite:

• Python 3.7+
• PyTorch 1.12+ and corresponding torchvision

Clone our repository:

git clone https://github.com/hkchengrex/Tracking-Anything-with-DEVA.git

Install with pip:

cd Tracking-Anything-with-DEVA
pip install -e .

(If you encounter the File "setup.py" not found error, upgrade your pip with pip install --upgrade pip)

Download the pretrained models:

bash scripts/download_models.sh

Required for the text-prompted/automatic demo:

Install our fork of Grounded-Segment-Anything. Follow its instructions.

Grounding DINO installation might fail silently. Try python -c "from groundingdino.util.inference import Model as GroundingDINOModel". If you get a warning about running on CPU mode only, make sure you have CUDA_HOME set during Grounding DINO installation.

(Optional) For fast integer program solving in the semi-online setting:

Get your gurobi licence which is free for academic use. If a license is not found, we fall back to using PuLP which is slower and is not rigorously tested by us. All experiments are conducted with gurobi.

DEMO.md contains more details on the input arguments and tips on speeding up inference. You can always look at deva/inference/eval_args.py and deva/ext/ext_eval_args.py for a full list of arguments.

With gradio:

python demo/demo_gradio.py

Then visit the link that popped up on the terminal. If executing on a remote server, try port forwarding.

We have prepared an example in example/vipseg/12_1mWNahzcsAc (a clip from the VIPSeg dataset). The following two scripts segment the example clip using either Grounded Segment Anything with text prompts or SAM with automatic (points in grid) prompting.

Script (text-prompted):

python demo/demo_with_text.py --chunk_size 4 \
--img_path ./example/vipseg/images/12_1mWNahzcsAc \ 
--amp --temporal_setting semionline \
--size 480 \
--output ./example/output --prompt person.hat.horse

Script (automatic):

python demo/demo_automatic.py --chunk_size 4 \
--img_path ./example/vipseg/images/12_1mWNahzcsAc \ 
--amp --temporal_setting semionline \
--size 480 \
--output ./example/output

1. Running DEVA with your own detection model.
2. Running DEVA with detections to reproduce the benchmark results.
3. Training the DEVA model.

• On closed-set data, DEVA most likely does not work as well as end-to-end approaches. Joint training is (for now) still a better idea when you have enough target data.
• Positive detections are amplified temporally due to propagation. Having a detector with a lower false positive rate (i.e., a higher threshold) helps.
• If new objects are coming in and out all the time (e.g., in driving scenes), we will keep a lot of objects in the memory bank which unfortunately increases the false positive rate. Decreasing max_missed_detection_count might help since we delete objects from memory more eagerly.

@inproceedings{cheng2023tracking,
  title={Tracking Anything with Decoupled Video Segmentation},
  author={Cheng, Ho Kei and Oh, Seoung Wug and Price, Brian and Schwing, Alexander and Lee, Joon-Young},
  booktitle={ICCV},
  year={2023}
}

The demo would not be possible without ❤️ from the community:

Grounded Segment Anything: https://github.com/IDEA-Research/Grounded-Segment-Anything

Segment Anything: https://github.com/facebookresearch/segment-anything

XMem: https://github.com/hkchengrex/XMem

Title card generated with OpenPano: https://github.com/ppwwyyxx/OpenPano