Implementation of 《DenseSAM: Semantic Enhance SAM For Efficient Dense Object Segmentation》(IJCAI 2025)

Description

DenseSAM replaces location-based prompts with semantic guidance for automatic segmentation in dense scenarios, achieving remarkable performance with only ~3M training parameters. It offers a cost-effective and efficient solution for adapting SAM to dense object segmentation.

Installation

Create conda env

conda create -n densesam python=3.10
conda activate densesam

Clone codes and install dependences

git clone https://github.com/imAzhou/DenseSAM.git
cd DenseSAM
pip install -r requirements.txt && pip install -e .

Dataset Preparation

Dataset Download

You can access detailed information and download the corresponding datasets via the links provided in the table below.

Cell Datasets	Building Datasets
CoNIC [paper download link]	WHU [paper download link]
CPM17 [paper download link]	Inria[paper download link]
MoNuSeg [paper download link]	Massachusetts[paper download link]

Dataset Process

All datasets have been processed into a unified format. Please refer to the processing scripts for each dataset located in the process_datasets/ directory:

# For semantic dataset which no need boundary,like whu,inria and massachusetts.
Step 1: python process_datasets/whu/init_ann.py
Step 2: python process_datasets/whu/gene_coco_panoptic.py

# For instance dataset which need boundary,like conic,cpm17 and monuseg.
Step 1: python process_datasets/conic/create_init_ann.py
Step 2: python process_datasets/conic/gene_coco_panoptic.py
Step 3: python process_datasets/conic/gene_boundary_map.py

Minor differences may exist in the processing of different datasets. For details, please refer to the preprocessing code in each subdirectory under the process_datasets/ directory.

After configuring the path to your dataset in config/datasets/dataset_name.py, you can use the following script to preview the dataset and its annotations to verify that the preprocessing has been performed correctly:

python process_datasets/browse_dataset.py \
    configs/datasets/conic.py \
    visual_results/gt_visual/conic \
    --vis_nums 5 \
    --show_bbox

Usage

Training

python scripts/train.py \
    densesam/configs/datasets/conic.py \
    --record_save_dir logs/conic \
    --print_interval 20 \
    --device cuda:1

Evaluation

python scripts/test.py \
    logs/conic/2025_01_21_04_31_16/config.py \
    logs/conic/2025_01_21_04_31_16 \
    logs/conic/2025_01_21_04_31_16/checkpoints/best.pth \
    --visual_pred \
    --draw_func draw_cell_color \
    --visual_interval 1 \
    --seed 666

Inference

python scripts/inference_nometric.py \
    logs/conic/2025_01_21_04_31_16/config.py \
    logs/conic/2025_01_21_04_31_16 \
    logs/conic/2025_01_21_04_31_16/checkpoints/best.pth \
    imgset_dir \
    --visual_interval 1

Contact

If you have any questions, please contact us via email: zhoulyaxx@zju.edu.cn