SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining (ICCV 2023)

plot

This is the official repository for the work [SparseDet: Improving Sparsely Annotated Object Detection with Pseudo-positive Mining] accepted to ICCV 2023. It includes scripts to train SparseDet on different splits also provided in this repository. Also see our Project Webpage.

Setup

Tested with Python 3.6.15

Create Environment

python3 -m venv env_sparsedet

Activate Environment and Install Packages

source env_sparsedet/bin/activate
pip install --upgrade pip
pip install torch==1.10.1+cu111 torchvision==0.11.2+cu111 torchaudio==0.10.1 -f https://download.pytorch.org/whl/cu111/torch_stable.html
python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.10/index.html
pip install opencv-python==4.6.0.66
pip install setuptools==59.5.0

Splits

Link to Splits - https://drive.google.com/drive/folders/168agXPO7LmpMWItl2bonbsdcEulYD2Cq?usp=sharing. Download them and place them in the splits directory.

Sample Commands for Training on 4 Gpus

#COCO
python plain_train_net.py \
--config-file configs/faster_rcnn_R_101_FPN_3x_mod.yaml \
--dist-url tcp://0.0.0.0:12345 \
--num-gpus 4 \
--resume \
OUTPUT_DIR experiments_coco_fpn/split1_30p \
DATASETS.TRAIN split1_30p \
DATASETS.TEST coco_val \
DATALOADER.NUM_WORKERS 8 \
SOLVER.IMS_PER_BATCH 8 \
FIXMATCH True \
FIXMATCH_STRONG_AUG True \
MASK_BOXES 30000 \
MASK_BOXES_THRESH 0.8 \
MASK_BOXES_RPN True \
DISTILLATION_LOSS_WEIGHT 1.0 \
DET_THRESH 0.0 \
CONSISTENCY_REGULARIZATION False \
MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE 256 \
MODEL.ROI_HEADS.POSITIVE_FRACTION 0.5 \
SOLVER.BASE_LR 0.01 \
SEED 1234

#VOC
python plain_train_net.py \
--config-file configs/faster_rcnn_R_101_FPN_3x_voc_18k.yaml \
--dist-url tcp://0.0.0.0:12345 \
--num-gpus 4 \
--resume \
OUTPUT_DIR experiments_voc_fpn/split5_50p \
DATASETS.TRAIN split5_50p \
DATASETS.TEST voc_test \
DATALOADER.NUM_WORKERS 8 \
SOLVER.IMS_PER_BATCH 8 \
FIXMATCH True \
FIXMATCH_STRONG_AUG True \
MASK_BOXES 9000 \
MASK_BOXES_THRESH 0.8 \
MASK_BOXES_RPN True \
DISTILLATION_LOSS_WEIGHT 1.0 \
DET_THRESH 0.0 \
CONSISTENCY_REGULARIZATION False \
MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE 256 \
MODEL.ROI_HEADS.POSITIVE_FRACTION 0.5 \
SOLVER.BASE_LR 0.01 \
SEED 1234

License

Distributed under the MIT License.

saksham-s / SparseDet