Code Implementation for CVPR2024 PAPER -- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (PnP-OVSS)

❗ This is the updated verision of PnP-OVSS with the bug corrected. Please refer to our latest arxiv version of the paper for the updated results and procedures. Only the code for PnP-OVSS + BLIP is provided here. Will update BridgeTower shortly.

System

CUDA Version: 11.7
GPU Memory: 48GB

Download LAVIS

Build LAVIS environment following the instruction here

conda create -n lavis python=3.8 
conda activate lavis 
pip install salesforce-lavis 
git clone https://github.com/salesforce/LAVIS.git 
cd LAVIS 
pip install -e .

This would download the newest torch, may need to modify torch version based on your cuda version.
Might also need to downgrade transformer pip install transformers==4.25

Download Gradient_Free_Optimizers_master

Download Gradient_Free_Optimizers_master and put it under LAVIS (This is for Random Search. Can ignore it for now)

Download pydensecrf

Git clone pydensecrf and put it under LAVIS

Download datasets

ADE20K can be downloaded from here
COCO Object can be downloaded from here
May Download the following datasets following instruction from mmsegmentation
Pascal VOC
Pascal Context
COCO Stuff

LAVIS
├──mmsegmentation
│   ├── mmseg
│   ├── tools
│   ├── configs
│   ├── data
│   │   ├── VOCdevkit
│   │   │   ├── VOC2012
│   │   │   │   ├── JPEGImages
│   │   │   │   ├── SegmentationClass
│   │   │   │   ├── Annotations
│   │   │   ├── VOC2010
│   │   │   │   ├── JPEGImages
│   │   │   │   ├── SegmentationClassContext
│   │   │   │   ├── ImageSets
│   │   │   │   │   ├── SegmentationContext
│   │   │   │   │   │   ├── train.txt
│   │   │   │   │   │   ├── val.txt
│   │   │   │   ├── trainval_merged.json
│   │   ├── coco_stuff164k
│   │   │   ├── images
│   │   │   │   ├── train2017
│   │   │   │   ├── val2017
│   │   │   ├── annotations
│   │   │   │   ├── train2017
│   │   │   │   ├── val2017
├── coco
│   ├── images
│   │   ├── val2017
├── ADEChallengeData2016
│   ├── images
│   │   ├── validation
│   ├── annotations
│   │   ├── validation

Replace the config file and GradCAM file of BLIP

Download all the files in this repository and put them under LAVIS
Replace /home/user/LAVIS/lavis/models/blip_models/blip_image_text_matching.py with the file in this repository
Replace /home/user/LAVIS/lavis/configs/models/blip_itm_large.yaml with the file in this repository
Replace /home/user/LAVIS/lavis/models/med.py with the file in this repository
Replace /home/user/LAVIS/lavis/models/vit.py with the file in this repository
Replace /home/user/LAVIS/lavis/models/base_model.py with the file in this repository
Replace /home/user/LAVIS/lavis/processors/blip_processors.py with the file in this repository

Obtain the GPT4o Classification

We provide the classification result from GPT4o for the five datasets in this repository, if you would like to run GPT4o by yourself, you may obtain your GPT KPI Key and run:

python PnP_OVSS_0514_updated_segmentation.py \
  --apikey xxxx
  --output_dir GPT4o_classification
  --data_type coco_object 

Run scripts

For Pascal VOC, Pascal Context, and ADE20K, modify the data_type to "voc", "psc", and "ade20k"

You may modify world_size to run on multiple GPUs

Run bash Run_seg.sh

python PnP_OVSS_0514_updated_segmentation.py \
  --home_dir \home\letitiabanana\LAVIS \
  --save_path BLIP_0602_336_ADE20K_segmentation\
  --master_port 29790 --gen_multiplecap_withpnpvqa label --world_size 1 \
  --img_size 336 \
  --del_patch_num sort_thresh005 \
  --batch_size 35 \
  --max_att_block_num 8 --drop_iter 4 --prune_att_head 9 --sort_threshold 0.05 \
  --threshold 0.15 \
  --postprocess blur+crf \
  --data_type ade20k

Then modify the save_path and data_type in "Calculate_mIoU.sh" and run bash calculate_mIou.sh to get the final result for the datasets

python Calculate_mIoU.py \
--save_path BLIP_0602_336_ADE20K_segmentation \
--data_type ade20k

For COCO Object and COCO Stuff, set the data_typa argument in the bash file as "coco_object" or "coco_stuff"

run bash Run_seg_coco.sh

python PnP_OVSS_0514_updated_segmentation_coco.py \
  --home_dir \home\letitiabanana\LAVIS \
  --save_path BLIP_0602_336_ADE20K_segmentation\
  --master_port 29790 --gen_multiplecap_withpnpvqa label --world_size 1 \
  --img_size 336 \
  --del_patch_num sort_thresh005 \
  --batch_size 35 \
  --max_att_block_num 8 --drop_iter 4 --prune_att_head 9 --sort_threshold 0.05 \
  --threshold 0.15 \
  --postprocess blur+crf \
  --data_type coco_object

Then modify the save_path and data_type in "Calculate_mIoU.sh" and run bash calculate_mIou.sh to get the final result for the datasets

For modifying image size, you may also need to modify the image size in /home/user/LAVIS/lavis/configs/models/blip_itm_large.yaml

Please cite us with this BibTeX:

@article{luo2023emergent,
  title={Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models},
  author={Luo, Jiayun and Khandelwal, Siddhesh and Sigal, Leonid and Li, Boyang},
  journal={arXiv e-prints 2311.17095},
  year={2023}
}