Home

Awesome

Grounded SAM 2: Ground and Track Anything in Videos

IDEA-Research

Tianhe Ren, Shuo Shen

[SAM 2 Paper] [Grounding DINO Paper] [Grounding DINO 1.5 Paper] [DINO-X Paper] [BibTeX]

Video Name

Highlights

Grounded SAM 2 is a foundation model pipeline towards grounding and track anything in Videos with Grounding DINO, Grounding DINO 1.5, Florence-2, DINO-X and SAM 2.

In this repo, we've supported the following demo with simple implementations:

Grounded SAM 2 does not introduce significant methodological changes compared to Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks. Both approaches leverage the capabilities of open-world models to address complex visual tasks. Consequently, we try to simplify the code implementation in this repository, aiming to enhance user convenience.

Latest updates

Contents

Installation

Download the pretrained SAM 2 checkpoints:

cd checkpoints
bash download_ckpts.sh

Download the pretrained Grounding DINO checkpoints:

cd gdino_checkpoints
bash download_ckpts.sh

Installation without docker

Install PyTorch environment first. We use python=3.10, as well as torch >= 2.3.1, torchvision>=0.18.1 and cuda-12.1 in our environment to run this demo. Please follow the instructions here to install both PyTorch and TorchVision dependencies. Installing both PyTorch and TorchVision with CUDA support is strongly recommended. You can easily install the latest version of PyTorch as follows:

pip3 install torch torchvision torchaudio

Since we need the CUDA compilation environment to compile the Deformable Attention operator used in Grounding DINO, we need to check whether the CUDA environment variables have been set correctly (which you can refer to Grounding DINO Installation for more details). You can set the environment variable manually as follows if you want to build a local GPU environment for Grounding DINO to run Grounded SAM 2:

export CUDA_HOME=/path/to/cuda-12.1/

Install Segment Anything 2:

pip install -e .

Install Grounding DINO:

pip install --no-build-isolation -e grounding_dino

Installation with docker

Build the Docker image and Run the Docker container:

cd Grounded-SAM-2
make build-image
make run

After executing these commands, you will be inside the Docker environment. The working directory within the container is set to: /home/appuser/Grounded-SAM-2

Once inside the Docker environment, you can start the demo by running:

python grounded_sam2_tracking_demo.py

Grounded SAM 2 Demos

Grounded SAM 2 Image Demo (with Grounding DINO)

Note that Grounding DINO has already been supported in Huggingface, so we provide two choices for running Grounded SAM 2 model:

python grounded_sam2_hf_model_demo.py

[!NOTE] 🚨 If you encounter network issues while using the HuggingFace model, you can resolve them by setting the appropriate mirror source as export HF_ENDPOINT=https://hf-mirror.com

python grounded_sam2_local_demo.py

Grounded SAM 2 Image Demo (with Grounding DINO 1.5 & 1.6)

We've already released our most capable open-set detection model Grounding DINO 1.5 & 1.6, which can be combined with SAM 2 for stronger open-set detection and segmentation capability. You can apply the API token first and run Grounded SAM 2 with Grounding DINO 1.5 as follows:

Install the latest DDS cloudapi:

pip install dds-cloudapi-sdk --upgrade

Apply your API token from our official website here: request API token.

python grounded_sam2_gd1.5_demo.py

SAHI (Slicing Aided Hyper Inference) with Grounding DINO 1.5 and SAM 2

If your images are high resolution with dense objects, directly using Grounding DINO 1.5 for inference on the original image may not be the best choice. We support SAHI (Slicing Aided Hyper Inference), which works by first dividing the original image into smaller overlapping patches. Inference is then performed separately on each patch, and the final detection results are merged. This method is highly effective and accuracy for dense and small objects detection in high resolution images.

You can run SAHI inference by setting the following param in grounded_sam2_gd1.5_demo.py:

WITH_SLICE_INFERENCE = True

The visualization is shown as follows:

Text PromptInput ImageGrounded SAM 2Grounded SAM 2 with SAHI
Person

Grounded SAM 2 Image Demo (with DINO-X)

We've implemented Grounded SAM 2 with the strongest open-world perception model DINO-X for better open-set detection and segmentation performance. You can apply the API token first and run Grounded SAM 2 with DINO-X as follows:

Install the latest DDS cloudapi:

pip install dds-cloudapi-sdk --upgrade

Apply your API token from our official website here: request API token.

python grounded_sam2_dinox_demo.py

Automatically Saving Grounding Results (Image Demo)

After setting DUMP_JSON_RESULTS=True in the following Grounded SAM 2 Image Demos:

The grounding and segmentation results will be automatically saved in the outputs dir with the following format:

{
    "image_path": "path/to/image.jpg",
    "annotations": [
        {
            "class_name": "class_name",
            "bbox": [x1, y1, x2, y2],
            "segmentation": {
                "size": [h, w],
                "counts": "rle_encoded_mask"
            },
            "score": confidence score
        }
    ],
    "box_format": "xyxy",
    "img_width": w,
    "img_height": h
}

Grounded SAM 2 Video Object Tracking Demo

Based on the strong tracking capability of SAM 2, we can combined it with Grounding DINO for open-set object segmentation and tracking. You can run the following scripts to get the tracking results with Grounded SAM 2:

python grounded_sam2_tracking_demo.py

Support Various Prompt Type for Tracking

We've supported different types of prompt for Grounded SAM 2 tracking demo:

Grounded SAM 2 Tracking Pipeline

Grounded SAM 2 Video Object Tracking Demo (with Grounding DINO 1.5 & 1.6)

We've also support video object tracking demo based on our stronger Grounding DINO 1.5 model and SAM 2, you can try the following demo after applying the API keys for running Grounding DINO 1.5:

python grounded_sam2_tracking_demo_with_gd1.5.py

Grounded SAM 2 Video Object Tracking Demo with Custom Video Input (with Grounding DINO)

Users can upload their own video file (e.g. assets/hippopotamus.mp4) and specify their custom text prompts for grounding and tracking with Grounding DINO and SAM 2 by using the following scripts:

python grounded_sam2_tracking_demo_custom_video_input_gd1.0_hf_model.py

If you are not convenient to use huggingface demo, you can also run tracking demo with local grounding dino model with the following scripts:

python grounded_sam2_tracking_demo_custom_video_input_gd1.0_local_model.py

Grounded SAM 2 Video Object Tracking Demo with Custom Video Input (with Grounding DINO 1.5 & 1.6)

Users can upload their own video file (e.g. assets/hippopotamus.mp4) and specify their custom text prompts for grounding and tracking with Grounding DINO 1.5 and SAM 2 by using the following scripts:

python grounded_sam2_tracking_demo_custom_video_input_gd1.5.py

You can specify the params in this file:

VIDEO_PATH = "./assets/hippopotamus.mp4"
TEXT_PROMPT = "hippopotamus."
OUTPUT_VIDEO_PATH = "./hippopotamus_tracking_demo.mp4"
API_TOKEN_FOR_GD1_5 = "Your API token" # api token for G-DINO 1.5
PROMPT_TYPE_FOR_VIDEO = "mask" # using SAM 2 mask prediction as prompt for video predictor

After running our demo code, you can get the tracking results as follows:

Video Name

And we will automatically save the tracking visualization results in OUTPUT_VIDEO_PATH.

[!WARNING] We initialize the box prompts on the first frame of the input video. If you want to start from different frame, you can refine ann_frame_idx by yourself in our code.

Grounded SAM 2 Video Object Tracking Demo with Custom Video Input (with DINO-X)

Users can upload their own video file (e.g. assets/hippopotamus.mp4) and specify their custom text prompts for grounding and tracking with DINO-X and SAM 2 by using the following scripts:

python grounded_sam2_tracking_demo_custom_video_input_dinox.py

Grounded-SAM-2 Video Object Tracking with Continuous ID (with Grounding DINO)

In above demos, we only prompt Grounded SAM 2 in specific frame, which may not be friendly to find new object during the whole video. In this demo, we try to find new objects and assign them with new ID across the whole video, this function is still under develop. it's not that stable now.

Users can upload their own video files and specify custom text prompts for grounding and tracking using the Grounding DINO and SAM 2 frameworks. To do this, execute the script:

python grounded_sam2_tracking_demo_with_continuous_id.py

You can customize various parameters including:

After running our demo code, you can get the tracking results as follows:

Video Name

If you want to try Grounding DINO 1.5 model, you can run the following scripts after setting your API token:

python grounded_sam2_tracking_demo_with_continuous_id_gd1.5.py

Grounded-SAM-2 Video Object Tracking with Continuous ID plus Reverse Tracking(with Grounding DINO)

This method could simply cover the whole lifetime of the object

python grounded_sam2_tracking_demo_with_continuous_id_plus.py

Grounded SAM 2 Florence-2 Demos

Grounded SAM 2 Florence-2 Image Demo

In this section, we will explore how to integrate the feature-rich and robust open-source models Florence-2 and SAM 2 to develop practical applications.

Florence-2 is a powerful vision foundation model by Microsoft which supports a series of vision tasks by prompting with special task_prompt includes but not limited to:

TaskTask PromptText InputTask Introduction
Object Detection<OD>✘Detect main objects with single category name
Dense Region Caption<DENSE_REGION_CAPTION>✘Detect main objects with short description
Region Proposal<REGION_PROPOSAL>✘Generate proposals without category name
Phrase Grounding<CAPTION_TO_PHRASE_GROUNDING>✔Ground main objects in image mentioned in caption
Referring Expression Segmentation<REFERRING_EXPRESSION_SEGMENTATION>✔Ground the object which is most related to the text input
Open Vocabulary Detection and Segmentation<OPEN_VOCABULARY_DETECTION>✔Ground any object with text input

Integrate Florence-2 with SAM-2, we can build a strong vision pipeline to solve complex vision tasks, you can try the following scripts to run the demo:

[!NOTE] 🚨 If you encounter network issues while using the HuggingFace model, you can resolve them by setting the appropriate mirror source as export HF_ENDPOINT=https://hf-mirror.com

Object Detection and Segmentation

python grounded_sam2_florence2_image_demo.py \
    --pipeline object_detection_segmentation \
    --image_path ./notebooks/images/cars.jpg

Dense Region Caption and Segmentation

python grounded_sam2_florence2_image_demo.py \
    --pipeline dense_region_caption_segmentation \
    --image_path ./notebooks/images/cars.jpg

Region Proposal and Segmentation

python grounded_sam2_florence2_image_demo.py \
    --pipeline region_proposal_segmentation \
    --image_path ./notebooks/images/cars.jpg

Phrase Grounding and Segmentation

python grounded_sam2_florence2_image_demo.py \
    --pipeline phrase_grounding_segmentation \
    --image_path ./notebooks/images/cars.jpg \
    --text_input "The image shows two vintage Chevrolet cars parked side by side, with one being a red convertible and the other a pink sedan, \
            set against the backdrop of an urban area with a multi-story building and trees. \
            The cars have Cuban license plates, indicating a location likely in Cuba."

Referring Expression Segmentation

python grounded_sam2_florence2_image_demo.py \
    --pipeline referring_expression_segmentation \
    --image_path ./notebooks/images/cars.jpg \
    --text_input "The left red car."

Open-Vocabulary Detection and Segmentation

python grounded_sam2_florence2_image_demo.py \
    --pipeline open_vocabulary_detection_segmentation \
    --image_path ./notebooks/images/cars.jpg \
    --text_input "car <and> building"

Grounded SAM 2 Florence-2 Image Auto-Labeling Demo

Florence-2 can be used as a auto image annotator by cascading its caption capability with its grounding capability.

TaskTask PromptText Input
Caption + Phrase Grounding<CAPTION> + <CAPTION_TO_PHRASE_GROUNDING>✘
Detailed Caption + Phrase Grounding<DETAILED_CAPTION> + <CAPTION_TO_PHRASE_GROUNDING>✘
More Detailed Caption + Phrase Grounding<MORE_DETAILED_CAPTION> + <CAPTION_TO_PHRASE_GROUNDING>✘

You can try the following scripts to run these demo:

Caption to Phrase Grounding

python grounded_sam2_florence2_autolabel_pipeline.py \
    --image_path ./notebooks/images/groceries.jpg \
    --pipeline caption_to_phrase_grounding \
    --caption_type caption

Citation

If you find this project helpful for your research, please consider citing the following BibTeX entry.

@misc{ravi2024sam2segmentimages,
      title={SAM 2: Segment Anything in Images and Videos}, 
      author={Nikhila Ravi and Valentin Gabeur and Yuan-Ting Hu and Ronghang Hu and Chaitanya Ryali and Tengyu Ma and Haitham Khedr and Roman Rädle and Chloe Rolland and Laura Gustafson and Eric Mintun and Junting Pan and Kalyan Vasudev Alwala and Nicolas Carion and Chao-Yuan Wu and Ross Girshick and Piotr Dollár and Christoph Feichtenhofer},
      year={2024},
      eprint={2408.00714},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.00714}, 
}

@article{liu2023grounding,
  title={Grounding dino: Marrying dino with grounded pre-training for open-set object detection},
  author={Liu, Shilong and Zeng, Zhaoyang and Ren, Tianhe and Li, Feng and Zhang, Hao and Yang, Jie and Li, Chunyuan and Yang, Jianwei and Su, Hang and Zhu, Jun and others},
  journal={arXiv preprint arXiv:2303.05499},
  year={2023}
}

@misc{ren2024grounding,
      title={Grounding DINO 1.5: Advance the "Edge" of Open-Set Object Detection}, 
      author={Tianhe Ren and Qing Jiang and Shilong Liu and Zhaoyang Zeng and Wenlong Liu and Han Gao and Hongjie Huang and Zhengyu Ma and Xiaoke Jiang and Yihao Chen and Yuda Xiong and Hao Zhang and Feng Li and Peijun Tang and Kent Yu and Lei Zhang},
      year={2024},
      eprint={2405.10300},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@misc{ren2024grounded,
      title={Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks}, 
      author={Tianhe Ren and Shilong Liu and Ailing Zeng and Jing Lin and Kunchang Li and He Cao and Jiayu Chen and Xinyu Huang and Yukang Chen and Feng Yan and Zhaoyang Zeng and Hao Zhang and Feng Li and Jie Yang and Hongyang Li and Qing Jiang and Lei Zhang},
      year={2024},
      eprint={2401.14159},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@article{kirillov2023segany,
  title={Segment Anything}, 
  author={Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and Mao, Hanzi and Rolland, Chloe and Gustafson, Laura and Xiao, Tete and Whitehead, Spencer and Berg, Alexander C. and Lo, Wan-Yen and Doll{\'a}r, Piotr and Girshick, Ross},
  journal={arXiv:2304.02643},
  year={2023}
}

@misc{jiang2024trex2,
      title={T-Rex2: Towards Generic Object Detection via Text-Visual Prompt Synergy}, 
      author={Qing Jiang and Feng Li and Zhaoyang Zeng and Tianhe Ren and Shilong Liu and Lei Zhang},
      year={2024},
      eprint={2403.14610},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}