JDet-WOOD

Unified Jittor implementations for weakly supervised oriented object detection

JDet-WOOD brings six related weakly supervised oriented object detection methods into one installable JDet codebase. All models share the same python/jdet package, runner, datasets, rotated operators, and command-line interface; select a method by changing only the config file.

Supported models

Method	Supervision	JDet model type	Primary config
H2RBox	horizontal boxes	H2RBox	h2rbox_obb_r50_adamw_fpn_1x_dota.py
H2RBox-v2	horizontal boxes	H2RBoxV2P	h2rbox_v2p_obb_r50_adamw_fpn_1x_dota.py
Wholly-WOOD	points, HBoxes, RBoxes, or mixed labels	WhollyWood	whollywood_obb_r50_adamw_fpn_1x_dota.py
Point2RBox	points	Point2RBox	point2rbox_obb_r50_adamw_fpn_1x_dota.py
Point2RBox-v2	points	Point2RBoxV2	point2rbox_v2_final_fixed.py
Point2RBox-v3	points	Point2RBoxV3	point2rbox_v3_1x_dota.py

The Wholly-WOOD family lives under configs/whollywood, while the newer Point2RBox releases keep their stage-1, pseudo-label, and stage-2 configs under configs/point2rbox_v2 and configs/point2rbox_v3.

Installation

Validated environment:

• Linux, Python 3.10
• Jittor 1.3.8.5
• NumPy 1.26.4 (NumPy 2.x is not supported by this Jittor release)
• CUDA 11.2 and g++-10 for the validated GPU setup

git clone https://github.com/VisionXLab/JDet-WOOD.git
cd JDet-WOOD
bash scripts/setup_env.sh
conda activate jdet-wood
export PYTHONPATH="$PWD:$PWD/python"

For an existing compatible environment:

python -m pip install -r requirements.txt
python -m pip install -e .
export cc_path=/usr/bin/g++-10
export PYTHONPATH="$PWD:$PWD/python"

See docs/environment.md for compiler, CUDA, and Jittor compatibility notes.

Data and auxiliary weights

The released DOTA-v1.0 configs use 1024×1024 patches with a 200-pixel gap and expect the validated split under /root/data/split_ss_dota:

/root/data/split_ss_dota/
├── trainval/
│   ├── images/
│   └── annfiles/
└── test/
    └── images/

Update the dataset paths in the selected config if your data is elsewhere. The JDet preprocessing utilities and additional supported datasets are documented in docs/data.md and docs/dota.md.

Point2RBox-v3 additionally expects converted MobileSAM and TED weights:

weights/mobile_sam.pkl
weights/ted.pkl

Convert original PyTorch weights when needed:

python tools/convert_torch_weights.py /path/to/mobile_sam.pt weights/mobile_sam.pkl
python tools/convert_ted_weights.py /path/to/ted.pth weights/ted.pkl

Converted weights, checkpoints, and logs are available from the Point2RBox-v3 Jittor collection and the Point2RBox-v2 Jittor collection.

Train and test

Every model uses the same entry point:

python tools/run_net.py --config-file <config> --task train
python tools/run_net.py --config-file <config> --task test

Examples:

# Wholly-WOOD, H2RBox, H2RBox-v2, and Point2RBox
python tools/run_net.py --config-file configs/whollywood/whollywood_obb_r50_adamw_fpn_1x_dota.py --task train
python tools/run_net.py --config-file configs/whollywood/h2rbox_obb_r50_adamw_fpn_1x_dota.py --task train
python tools/run_net.py --config-file configs/whollywood/h2rbox_v2p_obb_r50_adamw_fpn_1x_dota.py --task train
python tools/run_net.py --config-file configs/whollywood/point2rbox_obb_r50_adamw_fpn_1x_dota.py --task train

# Point2RBox-v2 and Point2RBox-v3 stage 1
python tools/run_net.py --config-file configs/point2rbox_v2/point2rbox_v2_final_fixed.py --task train
python tools/run_net.py --config-file configs/point2rbox_v3/point2rbox_v3_1x_dota.py --task train

Set resume_path in the selected config to evaluate a downloaded checkpoint, then run with --task val or --task test.

Point2RBox-v2 two-stage workflow

# 1. Train the end-to-end point-supervised model.
CUDA_VISIBLE_DEVICES=0 python tools/run_net.py \
  --config-file configs/point2rbox_v2/point2rbox_v2_final_fixed.py \
  --task train

# 2. Export rotated pseudo labels.
CUDA_VISIBLE_DEVICES=0 python tools/generate_pseudo_labels.py \
  --config configs/point2rbox_v2/point2rbox_v2_pseudo_generator_dota.py \
  --ckpt work_dirs/point2rbox_v2_1x_dota_final_fixed/checkpoints/ckpt_12.pkl \
  --out /root/data/split_ss_dota/point2rbox_v2_pseudo_labels

# 3. Train rotated FCOS from the pseudo labels.
CUDA_VISIBLE_DEVICES=0 python tools/run_net.py \
  --config-file configs/point2rbox_v2/rotated_fcos_1x_dota_using_pseudo.py \
  --task train

tools/auto_stage2_pipeline.sh automates the same workflow and accepts GPU_ID, ENV_NAME, STAGE1_PID, and OUT_DIR environment overrides.

Point2RBox-v3 two-stage workflow

# 1. Train the end-to-end detector.
CUDA_VISIBLE_DEVICES=0 python tools/run_net.py \
  --config-file configs/point2rbox_v3/point2rbox_v3_1x_dota.py \
  --task train

# 2. Export SAM-refined pseudo labels.
CUDA_VISIBLE_DEVICES=0 python tools/export_pseudo_labels.py \
  --config-file configs/point2rbox_v3/point2rbox_v3_pseudo_generator_dota.py \
  --ckpt work_dirs/point2rbox_v3_1x_dota/checkpoints/ckpt_12.pkl

# 3. Train rotated FCOS from the pseudo labels.
CUDA_VISIBLE_DEVICES=0 python tools/run_net.py \
  --config-file configs/point2rbox_v3/rotated_fcos_1x_dota_using_pseudo.py \
  --task train

Reproduced Point2RBox results

DOTA-v1.0 Task1 mAP50 on the official test server:

Model	Paper	Jittor	Checkpoint
Point2RBox-v2 end-to-end	51.00	48.95	download
Point2RBox-v2 + rotated FCOS	62.61	59.39	download
Point2RBox-v3 end-to-end	59.61	59.52	download
Point2RBox-v3 + rotated FCOS	66.09	65.50	download

Verification

The test suite covers registry/config integration, Jittor numerics, rotated geometry and losses, optimizer resume, datasets, MobileSAM, TED, pseudo-label serialization, and v2/v3 detector routing.

export PYTHONPATH="$PWD:$PWD/python"
python -m pytest tests/test_jdet_wood_registry.py tests/smoke -q
python -m pytest tests/parity tests/test_v3_norm_eval.py -q
python tests/test_sam.py
python tests/test_ted.py

GPU-only parity tests require the validated CUDA environment. The repository does not bundle DOTA data or released model checkpoints.

Source snapshots

JDet-WOOD was unified from these VisionXLab repositories:

Source	Imported commit	Role
h2rbox-jittor	90e756ba375cfa74ad55bf57527ed17cb4d1ebbe	H2RBox provenance
whollywood-jittor	6bca5e07d5ea60ba2f22f06ea90961e0b4235b37	Wholly-WOOD family
Point2RBox-v2-jittor	66bf12fa6764c44da4046018159b1d7e56b9c249	v2 release fixes and pipeline
Point2RBox-v3-jittor	d309b47f060bae040f7889cfd187c2ffc393db5f	unified base and v3 implementation

Please cite the corresponding method papers when using a model. BibTeX entries for the Wholly-WOOD, H2RBox, H2RBox-v2, and Point2RBox family are collected in the Point2RBox-v2 source README.

Acknowledgements and license

Built on Jittor, JDet, MobileSAM, and the VisionXLab weakly supervised oriented detection projects listed above.

Released under the Apache License 2.0. See LICENSE.txt.