📷 EVF-SAM

Early Vision-Language Fusion for Text-Prompted Segment Anything Model

Yuxuan Zhang^1,*, Tianheng Cheng^1,*, Lei Liu², Heng Liu², Longjin Ran², Xiaoxin Chen², Wenyu Liu¹, Xinggang Wang^1,📧

¹ Huazhong University of Science and Technology, ² vivo AI Lab

(* equal contribution, 📧 corresponding author)

We have expanded our EVF-SAM to powerful SAM-2. Besides improvements on image prediction, our new model also performs well on video prediction (powered by SAM-2). Only at the expense of a simple image training process on RES datasets, we find our EVF-SAM has zero-shot video text-prompted capability. Try our code!

• EVF-SAM extends SAM's capabilities with text-prompted segmentation, achieving high accuracy in Referring Expression Segmentation.
• EVF-SAM is designed for efficient computation, enabling rapid inference in few seconds per image on a T4 GPU.

• Release code
• Release weights
• Release demo 👉 🤗 evf-sam
• Release code and weights based on SAM-2
• Update demo supporting SAM-2👉 🤗 evf-sam2
• release new checkpoint supporting body part segmentation and semantic level segmentation.
• update demo supporting multitask

Input text	Input image	Output
"zebra top left"
"a pizza with a yellow sign on top of it"
"the broccoli closest to the ketchup bottle"
"[semantic] hair"
"[semantic] sea"

1. Clone this repository
2. Install pytorch for your cuda version. Note that torch>=2.0.0 is needed if you are to use SAM-2, and torch>=2.2 is needed if you want to enable flash-attention. (We use torch==2.0.1 with CUDA 11.7 and it works fine.)
3. pip install -r requirements.txt
4. If you are to use the video prediction function, run:

cd model/segment_anything_2
python setup.py build_ext --inplace

1. -multimask checkpoints are only available with commits>=9d00853, while other checkpoints are available with commits<9d00853

2. -multimask checkpoints are jointly trained on Ref, ADE20k, Object365, PartImageNet, humanparsing, pascal part datasets. These checkpoints are able to segment part (e.g., hair, arm), background object (e.g., sky, ground), and semantic-level masks. (by adding special token "[semantic] " in front your prompt)

python inference.py  \
  --version <path to evf-sam> \
  --precision='fp16' \
  --vis_save_path "<path to your output direction>" \
  --model_type <"ori" or "effi" or "sam2", depending on your loaded ckpt>   \
  --image_path <path to your input image> \
  --prompt <customized text prompt>

--load_in_8bit and --load_in_4bit are optional
for example:

python inference.py  \
  --version YxZhang/evf-sam2 \
  --precision='fp16' \
  --vis_save_path "vis" \
  --model_type sam2   \
  --image_path "assets/zebra.jpg" \
  --prompt "zebra top left"

firstly slice video into frames

ffmpeg -i <your_video>.mp4 -q:v 2 -start_number 0 <frame_dir>/'%05d.jpg'

then:

python inference_video.py  \
  --version <path to evf-sam2> \
  --precision='fp16' \
  --vis_save_path "vis/" \
  --image_path <frame_dir>   \
  --prompt <customized text prompt>   \
  --model_type sam2

you can use frame2video.py to concat the predicted frames to a video.

image demo

python demo.py <path to evf-sam>

video demo

python demo_video.py <path to evf-sam2>

Referring segmentation datasets: refCOCO, refCOCO+, refCOCOg, refCLEF (saiapr_tc-12) and COCO2014train

├── dataset
│   ├── refer_seg
│   │   ├── images
│   │   |   ├── saiapr_tc-12 
│   │   |   └── mscoco
│   │   |       └── images
│   │   |           └── train2014
│   │   ├── refclef
│   │   ├── refcoco
│   │   ├── refcoco+
│   │   └── refcocog

torchrun --standalone --nproc_per_node <num_gpus> eval.py   \
    --version <path to evf-sam> \
    --dataset_dir <path to your data root>   \
    --val_dataset "refcoco|unc|val" \
    --model_type <"ori" or "effi" or "sam2", depending on your loaded ckpt>

We borrow some codes from LISA, unilm, SAM, EfficientSAM, SAM-2.

@article{zhang2024evfsamearlyvisionlanguagefusion,
      title={EVF-SAM: Early Vision-Language Fusion for Text-Prompted Segment Anything Model}, 
      author={Yuxuan Zhang and Tianheng Cheng and Rui Hu and Lei Liu and Heng Liu and Longjin Ran and Xiaoxin Chen and Wenyu Liu and Xinggang Wang},
      year={2024},
      eprint={2406.20076},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2406.20076}, 
}