leftthomas / IDIP

A PyTorch implementation of IDIP based on ECAI 2023 paper "Instance-aware Diffusion Implicit Process for Box-based Instance Segmentation"

IDIP

A PyTorch implementation of IDIP based on ECAI 2023 paper Instance-aware Diffusion Implicit Process for Box-based Instance Segmentation.

Requirements

• Anaconda
• PyTorch

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

• Detectron2

conda install tensorboard
conda install -c conda-forge pycocotools
pip install opencv-python-headless
pip install git+https://github.com/facebookresearch/detectron2.git

• MMDetection

pip install openmim
mim install mmcv 
mim install mmdet

Usage

Set th environment variable DETECTRON2_DATASETS to the directory where the dataset saved, for example: export DETECTRON2_DATASETS=/home/data. Then download the backbone weights from MEGA, put them in results.

To train the model with resnet50 backbone on COCO 2017 Train dataset:

python main.py --config-file configs/res50.yaml --num-gpus 2

Using tensorboard to visualize the training process:

tensorboard --logdir=results --bind_all

To evaluate the model with resnet50 backbone on COCO 2017 Val dataset:

python main.py --config-file configs/res50.yaml --eval-only MODEL.WEIGHTS results/model.pth

To visualize the results of a given image by using the pre-trained model:

python demo.py --config-file configs/res50.yaml --input image.jpg --output out.jpg --opts MODEL.WEIGHTS results/model.pth

Benchmarks

The models are trained on two NVIDIA Tesla V100-SXM2-32GB GPUs, and tested on COCO 2017 dataset. All the hyper-parameters are the default values.

Backbone	APVal	APTest	AP50	AP75	APS	APM	APL	Download
ResNet-50	39.5	40.2	63.0	43.7	22.2	43.3	53.3	MEGA
ResNet-101	40.4	41.9	65.1	45.5	22.8	45.2	55.7	MEGA

Results