This repository contains the code for the paper GSVA: Generalized Segmentation via Multimodal Large Language Models (CVPR2024). [arXiv][video][poster][CVPR page]
Generalized Referring Expression Segmentation (GRES) has recently been proposed to support multiple-target and empty-target cases, which are practical in real world application but omitted in classic RES tasks. GRES poses the challenges that require the model to handle complex spatial relationships of the instances in the image to segment the targets at various locations and reject the empty targets in the wrong places. Current segmentation MLLM focuses more on the complex referring expressions which could benefit from the LLM, however the multiple-target and empty-target challenges remain uncharted.
In this paper, we propose Generalized Segmentation Vision Assistant (GSVA) to address this gap. Specifically, GSVA reuses the [SEG] token to prompt the segmentation model towards supporting multiple mask references simultaneously and innovatively learns to generate a [REJ] token to reject the null targets explicitly.
The architecture of GSVA follows current segmentation MLLMs like LISA, which enables high-fidelity segmentation outputs by integrating two types of foundation models: (1) Multimodal Large Language Model (MLLM) as an aligned vision-language cognitive module; (2) Segmentation Foundation Model (SFM) to segment the target out of the input image based on user’s instruction. A special token [SEG] is used to prompt the SFM to follow the user instruction for desired referring segmentation outputs.
GSVA relax the one-to-one constraint to support multiple target outputs via learning multiple [SEG] tokens. By prompting each [SEG] token with the corresponding target expression, GSVA segments different mask for each target. For absent targets, GSVA predicts a [REJ] token to mark a rejection in the output sequence, therefore the empty target expressions are appropriately rejected.
More visualizations can be found in Appendix F of the paper.
bitsandbytes==0.43.3
deepspeed==0.15.0
einops==0.4.1
opencv-python-headless==4.10.0.84
peft==0.4.0
PyYAML==6.0.2
sentencepiece==0.2.0
termcolor==2.4.0
timm==0.6.13
tokenizers==0.19.1
torch==2.4.0+cu124
torchvision=0.19.0+cu124
transformers==4.44.2
We follow the instructions in LISA to prepare most of the datasets. For gRefCOCO, we add the expressions and annotations json files into the refer_seg sub-directory, as shown in the tree structure below.
├── ade20k
├── coco
├── cocostuff
│ ├── train2017
│ └── val2017
├── llava_dataset
│ └── llava_instruct_150k.json
├── mapillary
│ ├── LICENSE
│ ├── README
│ ├── config_v1.2.json
│ ├── config_v2.0.json
│ ├── demo.py
│ ├── testing
│ ├── training
│ └── validation
├── reason_seg
│ └── ReasonSeg
├── refer_seg
│ ├── grefcoco
│ ├── images
│ ├── refclef
│ ├── refcoco
│ ├── refcoco+
│ └── refcocog
└── vlpart
├── paco
└── pascal_part
We follow LISA with 3 variants of [LLaVA], for LLaVA-vicuna-v1.1, we apply the delta weights liuhaotian/LLaVA-Lightning-7B-delta-v1-1 and liuhaotian/LLaVA-13b-delta-v1-1 to llama-7b and llama-13b base weights, respectively. For LLaVA-Llama-2, we use the full weights of liuhaotian/llava-llama-2-13b-chat-lightning-preview. For LLaVA pretrained models, please see the instructions in LLaVA model zoo.
The pretrained weights of SAM can be found in Segment Anything.
The pretrained models can be found in OneDrive and TsinghuaCloud.
Modify the templates in scripts/eval_demo.sh with appropriate paths, then execute
bash scripts/eval_demo.sh
For example, set the weight argument in scripts/eval_demo.sh to gsva-7b-pt.bin, and gsva-7b-ft-gres.bin, then set the val_dataset to grefcoco|unc|val, grefcoco|unc|testA, and grefcoco|unc|testB, the evaluation scritps will give the following results:
# gsva-7b-pt.bin, grefcoco|unc|val, fp32 inference
[2024-09-08 08:04:56 default] (main.py 158): INFO Load trained weights successfully!
[2024-09-08 08:05:29 default] (main.py 205): INFO Testing with 1500 examples.
[2024-09-08 08:07:02 default] (solver.py 305): INFO [1] grefcoco giou: 0.6347, ciou: 0.6165, N_acc: 0.5698, T_acc: 0.9954.
# gsva-7b-pt.bin, grefcoco|unc|testA, fp32 inference
[2024-09-08 08:09:15 default] (main.py 158): INFO Load trained weights successfully!
[2024-09-08 08:09:47 default] (main.py 205): INFO Testing with 750 examples.
[2024-09-08 08:10:53 default] (solver.py 305): INFO [1] grefcoco giou: 0.6999, ciou: 0.6917, N_acc: 0.6339, T_acc: 0.9846.
# gsva-7b-pt.bin, grefcoco|unc|testB, fp32 inference
[2024-09-08 08:13:10 default] (main.py 158): INFO Load trained weights successfully!
[2024-09-08 08:13:41 default] (main.py 205): INFO Testing with 750 examples.
[2024-09-08 08:14:47 default] (solver.py 305): INFO [1] grefcoco giou: 0.6134, ciou: 0.6026, N_acc: 0.5868, T_acc: 0.9895.
# gsva-7b-ft-gres.bin, grefcoco|unc|val, fp32 inference
[2024-09-08 08:17:36 default] (main.py 158): INFO Load trained weights successfully!
[2024-09-08 08:18:08 default] (main.py 205): INFO Testing with 1500 examples.
[2024-09-08 08:19:40 default] (solver.py 305): INFO [1] grefcoco giou: 0.6643, ciou: 0.6319, N_acc: 0.6256, T_acc: 0.9962.
# gsva-7b-ft-gres.bin, grefcoco|unc|testA, fp32 inference
[2024-09-08 08:21:54 default] (main.py 158): INFO Load trained weights successfully!
[2024-09-08 08:22:25 default] (main.py 205): INFO Testing with 750 examples.
[2024-09-08 08:23:32 default] (solver.py 305): INFO [1] grefcoco giou: 0.7106, ciou: 0.6987, N_acc: 0.6572, T_acc: 0.9869.
# gsva-7b-ft-gres.bin, grefcoco|unc|testB, fp32 inference
[2024-09-08 08:25:46 default] (main.py 158): INFO Load trained weights successfully!
[2024-09-08 08:26:16 default] (main.py 205): INFO Testing with 750 examples.
[2024-09-08 08:27:19 default] (solver.py 305): INFO [1] grefcoco giou: 0.6220, ciou: 0.6035, N_acc: 0.6072, T_acc: 0.9910.
which corresponds to the Line 8 and Line 10 in Table 1 of the paper. Note that the numbers may vary w.r.t. the precision during inference and appear slightly differrent from them in the paper.
Modify the templates in scripts/train_demo.sh with appropriate paths, then execute
bash scripts/train_demo.sh
After the training process is done, use this file in the experiment working directory:
zero_to_fp32.py
to convert the ZeRO weight shards to a bin file to store the trained model weights. For training hyper-parameters, please refer to Appendix A of the paper.
If you want to start from a pretrained GSVA model for further finetuning, please consider set the weight argument to <path-to-your-model>.bin.
This code is developed on the top of LISA, LLaVA, SAM, and ReLA, we thank to their efficient and neat codebase. We also thank Dr. Yuan Yao and Prof. Zhiyuan Liu for their insightful and valuable comments on this research project. We also appreciate their generous support of computational resources.
If you find our work is useful in your research, please consider citing:
@InProceedings{Xia_2024_CVPR,
author = {Xia, Zhuofan and Han, Dongchen and Han, Yizeng and Pan, Xuran and Song, Shiji and Huang, Gao},
title = {GSVA: Generalized Segmentation via Multimodal Large Language Models},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2024},
pages = {3858-3869}
}
If you have any questions or concerns, please send email to xzf23@mails.tsinghua.edu.cn.
