small-object-detection-benchmark

🔥 our paper has been presented in ICIP 2022 Bordeaux, France (16-19 October 2022)

summary

small-object-detection benchmark on visdrone and xview datasets using fcos, vfnet and tood detectors

refer to Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection for full technical analysis

citation

If you use any file/result from this repo in your work, please cite it as:

@article{akyon2022sahi,
  title={Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection},
  author={Akyon, Fatih Cagatay and Altinuc, Sinan Onur and Temizel, Alptekin},
  journal={2022 IEEE International Conference on Image Processing (ICIP)},
  doi={10.1109/ICIP46576.2022.9897990},
  pages={966-970},
  year={2022}
}

visdrone results

refer to table 1 in Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection for more detail on visdrone results

setup	AP50	AP50s	AP50m	AP50l	results	checkpoints
FCOS+FI	25.8	14.2	39.6	45.1	download	request
FCOS+SAHI+PO	29.0	18.9	41.5	46.4	download	request
FCOS+SAHI+FI+PO	31.0	19.8	44.6	49.0	download	request
FCOS+SF+SAHI+PO	38.1	25.7	54.8	56.9	download	download
FCOS+SF+SAHI+FI+PO	38.5	25.9	55.4	59.8	download	download
---	---	---	---	---	---	---
VFNet+FI	28.8	16.8	44.0	47.5	download	request
VFNet+SAHI+PO	32.0	21.4	45.8	45.5	download	request
VFNet+SAHI+FI+PO	33.9	22.4	49.1	49.4	download	request
VFNet+SF+SAHI+PO	41.9	29.7	58.8	60.6	download	request
VFNet+SF+SAHI+FI+PO	42.2	29.6	59.2	63.3	download	request
---	---	---	---	---	---	---
TOOD+FI	29.4	18.1	44.1	50.0	download	request
TOOD+SAHI	31.9	22.6	44.0	45.2	download	request
TOOD+SAHI+PO	32.5	22.8	45.2	43.6	download	request
TOOD+SAHI+FI	34.6	23.8	48.5	53.1	download	request
TOOD+SAHI+FI+PO	34.7	23.8	48.9	50.3	download	request
TOOD+SF+FI	36.8	24.4	53.8	66.4	download	download
TOOD+SF+SAHI	42.5	31.6	58.0	61.1	download	download
TOOD+SF+SAHI+PO	43.1	31.7	59.0	60.2	download	download
TOOD+SF+SAHI+FI	43.4	31.7	59.6	65.6	download	download
TOOD+SF+SAHI+FI+PO	43.5	31.7	59.8	65.4	download	download

xview results

refer to table 2 in Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection for more detail on xview results

setup	AP50	AP50s	AP50m	AP50l	results	checkpoints
FCOS+FI	2.20	0.10	1.80	7.30	download	request
FCOS+SF+SAHI	15.8	11.9	18.4	11.0	download	download
FCOS+SF+SAHI+PO	17.1	12.2	20.2	12.8	download	download
FCOS+SF+SAHI+FI	15.7	11.9	18.4	14.3	download	download
FCOS+SF+SAHI+FI+PO	17.0	12.2	20.2	15.8	download	download
---	---	---	---	---	---	---
VFNet+FI	2.10	0.50	1.80	6.80	download	request
VFNet+SF+SAHI	16.0	11.9	17.6	13.1	download	download
VFNet+SF+SAHI+PO	17.7	13.7	19.7	15.4	download	download
VFNet+SF+SAHI+FI	15.8	11.9	17.5	15.2	download	download
VFNet+SF+SAHI+FI+PO	17.5	13.7	19.6	17.6	download	download
---	---	---	---	---	---	---
TOOD+FI	2.10	0.10	2.00	5.20	download	request
TOOD+SF+SAHI	19.4	14.6	22.5	14.2	download	download
TOOD+SF+SAHI+PO	20.6	14.9	23.6	17.0	download	download
TOOD+SF+SAHI+FI	19.2	14.6	22.3	14.7	download	download
TOOD+SF+SAHI+FI+PO	20.4	14.9	23.5	17.6	download	download

env setup

install pytorch:

conda install pytorch=1.10.0 torchvision=0.11.1 cudatoolkit=11.3 -c pytorch

install other requirements:

pip install -r requirements.txt

evaluation

• download desired checkpoint from the urls in readme.

• download xivew or visdrone dataset and convert to COCO format.

• set MODEL_PATH, MODEL_CONFIG_PATH, EVAL_IMAGES_FOLDER_DIR, EVAL_DATASET_JSON_PATH, INFERENCE_SETTING in predict_evaluate_analyse script then run the script.

roadmap

• add train test split support for xview to coco converter
• add mmdet config files (fcos, vfnet and tood) for xview training (9 train experiments)
• add mmdet config files (fcos, vfnet and tood) for visdrone training (9 train experiments)
• add coco result.json files, classwise coco eval results error analysis plots for all xview experiments
• add coco result.json files, classwise coco eval results error analysis plots for all visdrone experiments
• add .py scripts for inference + evaluation + error analysis using sahi