Official implementation of the paper "Semantic-SAM: Segment and Recognize Anything at Any Granularity"

Semantic-SAM: Segment and Recognize Anything at Any Granularity

In this work, we introduce Semantic-SAM, a universal image segmentation model to enable segment and recognize anything at any desired granularity. We have trained on the whole SA-1B dataset and our model can reproduce SAM and beyond it.

🍇 [Read our arXiv Paper]   🍎 [Try Gradio Demo1]   🍎 [Try Auto Generation Demo]

🚀 Features

🔥 Reproduce SAM. SAM training is a sub-task of ours. We have released the training code to reproduce SAM training.

🔥 Beyond SAM. Our newly proposed model offers the following attributes from instance to part level:

• Granularity Abundance. Our model can produce all possible segmentation granularities for a user click with high quality, which enables more controllable and user-friendly interactive segmentation.
• Semantic Awareness. We jointly train SA-1B with semantically labeled datasets to learn the semantics at both object-level and part-level.
• High Quality. We base on the DETR-based model to implement both generic and interactive segmentation, and validate that SA-1B helps generic and part segmentation. The mask quality of multi-granularity is high.

🚀 News

• We release the demo code for mask auto-generation!
• We release the demo code for interactive segmentation!
• We release the training and inference code and checkpoints (SwinT, SwinL) trained on SA-1B!
• We release the training code to reproduce SAM!

🔥 One-click to output up to 6 granularity masks. Try it in our demo! 🔥 Segment everything for one image. We output more masks with more granularity.

Our model supports a wide range of segmentation tasks and their related applications, including:

• Generic Segmentation
• Part Segmentation
• Interactive Multi-Granularity Segmentation with Semantics
• Multi-Granularity Image Editing

🔥 Related projects:

• Mask DINO: We build upon Mask DINO which is a unified detection and segmentation model to implement our model.
• OpenSeed: Strong open-set segmentation methods based on Mask DINO. We base on it to implement our open-vocabulary segmentation.
• SEEM: Segment using a wide range of user prompts.
• VLPart: Going denser with open-vocabulary part segmentation.

Getting Started

Installation

pip3 install torch==1.13.1 torchvision==0.14.1 --extra-index-url https://download.pytorch.org/whl/cu113
python -m pip install 'git+https://github.com/MaureenZOU/detectron2-xyz.git'
pip install git+https://github.com/cocodataset/panopticapi.git
git clone https://github.com/UX-Decoder/Semantic-SAM
cd Semantic-SAM
python -m pip install -r requirements.txt

export DATASET=/pth/to/dataset  # path to your coco data

Data preparation

Please refer to prepare SA-1B data. Let us know if you need more instructions about it.

Model Zoo

The currently released checkpoints are only trained with SA-1B data.

Demo

For interactive segmentation.

python demo.py

For mask auto-generation.

python demo_auto_generation.py

Evaluation

We do zero-shot evaluation on COCO val2017. $n is the number of gpus you use

For SwinL backbone

python train_net.py --eval_only --resume --num-gpus $n --config-file configs/semantic_sam_only_sa-1b_swinL.yaml COCO.TEST.BATCH_SIZE_TOTAL=$n  MODEL.WEIGHTS=/path/to/weights

For SwinT backbone

python train_net.py --eval_only --resume --num-gpus $n --config-file configs/semantic_sam_only_sa-1b_swinT.yaml COCO.TEST.BATCH_SIZE_TOTAL=$n  MODEL.WEIGHTS=/path/to/weights

Training

We currently release the code of training on SA-1B only. Complete training with semantics will be released later. $n is the number of gpus you use before running the training code, you need to specify your training data of SA-1B.

export SAM_DATASET=/pth/to/dataset
export SAM_DATASET_START=$start
export SAM_DATASET_END=$end

We convert SA-1B data into 100 tsv files. start(int, 0-99) is the start of your SA-1B data index and end(int, 0-99) is the end of your data index. If you are not using the tsv data formats, you can refer to this json registration for SAM for a reference.

For SwinL backbone

python train_net.py --resume --num-gpus $n  --config-file configs/semantic_sam_only_sa-1b_swinL.yaml COCO.TEST.BATCH_SIZE_TOTAL=$n  SAM.TEST.BATCH_SIZE_TOTAL=$n  SAM.TRAIN.BATCH_SIZE_TOTAL=$n MODEL.WEIGHTS=/path/to/weights

For SwinT backbone

python train_net.py --resume --num-gpus $n  --config-file configs/semantic_sam_only_sa-1b_swinT.yaml COCO.TEST.BATCH_SIZE_TOTAL=$n  SAM.TEST.BATCH_SIZE_TOTAL=$n  SAM.TRAIN.BATCH_SIZE_TOTAL=$n MODEL.WEIGHTS=/path/to/weights

We also support training to reproduce SAM

python train_net.py --resume --num-gpus $n  --config-file configs/semantic_sam_reproduce_sam_swinL.yaml COCO.TEST.BATCH_SIZE_TOTAL=$n  SAM.TEST.BATCH_SIZE_TOTAL=$n  SAM.TRAIN.BATCH_SIZE_TOTAL=$n MODEL.WEIGHTS=/path/to/weights

This is a swinL backbone. The only difference of this script is to use many-to-one matching and 3 prompts as in SAM.

Comparison with SAM and SA-1B Ground-truth

(a)(b) are the output masks of our model and SAM, respectively. The red points on the left-most image of each row are the user clicks. (c) shows the GT masks that contain the user clicks. The outputs of our model have been processed to remove duplicates.

Learned prompt semantics

We visualize the prediction of each content prompt embedding of points with a fixed order for our model. We find all the output masks are from small to large. This indicates each prompt embedding represents a semantic level. The red point in the first column is the click.

Method

Experiments

We also show that jointly training SA-1B interactive segmentation and generic segmentation can improve the generic segmentation performance.

We also outperform SAM on both mask quality and granularity completeness, please refer to our paper for more experimental details.