Patchmentation YOLOv5 🚀

This repository is forked from ultralytics/yolov5 and modified. For more information about the models, please visit the original repository.

This model YOLOv5 🚀 is used to benchmark patch augmentation performance of patchmentation.

Experiment Results and Comparison

Experiment	Dataset	Weights	P	R	mAP@.5	mAP@.5:.95
S-base	Pascal VOC 2007	yolov5s	0.685	0.558	0.586	0.327
S1	Pascal VOC 2007 with patch augmentation	yolov5s	0.715	0.621	0.671	0.405
S2	Pascal VOC 2007 with patch augmentation and soft-edge	yolov5s	0.717	0.624	0.67	0.403
S3	Pascal VOC 2007 with patch augmentation and negative-patch	yolov5s	0.726	0.607	0.665	0.393
S4	Pascal VOC 2007 with patch augmentation, soft-edge, and negative-patch	yolov5s	0.732	0.608	0.669	0.396
X-base	Pascal VOC 2007	yolov5x	0.81	0.688	0.745	0.516
X1	Pascal VOC 2007 with patch augmentation	yolov5x	0.817	0.719	0.776	0.556
X2	Pascal VOC 2007 with patch augmentation and soft-edge	yolov5x	0.804	0.704	0.772	0.544
X3	Pascal VOC 2007 with patch augmentation and negative-patch	yolov5x	0.796	0.718	0.776	0.549
X4	Pascal VOC 2007 with patch augmentation, soft-edge, and negative-patch	yolov5x	0.818	0.704	0.767	0.543
PS-base	Penn-Fudan-Ped	yolov5s	0.358	0.234	0.288	0.099
PS1	Single image from Campus - Garden1 with patch augmentation from Penn-Fudan-Ped	yolov5s	0.907	0.746	0.817	0.401
PX-base	Penn-Fudan-Ped	yolov5x	0.566	0.315	0.393	0.145
PX1	Single image from Campus - Garden1 with patch augmentation from Penn-Fudan-Ped	yolov5x	0.9	0.79	0.838	0.431

Dependency

• Using PIP

  pip install -r requirements.txt

• Using Docker (recommended)

  docker pull jstnxu/patchmentation:yolov5
  docker run -it --ipc=host --gpus all \
    -v {data_folder}:/patchmentation-dataset/data \
    -v {project_folder}:/workspace/runs/patchmentation \
    jstnxu/patchmentation:yolov5 /bin/bash

  • change {data_folder} to local path to load dataset.

  • change {project_folder} to local path to save outputs.

Arguments

Priority*	Arguments	Type	Description
-	--version	one or more str	Training version(s).
-	--overwrite	store_true	Overwrite existing output / zip.
-	--batch-size	int	Number of batch size. Required if train is true or test is true.
-	--epochs	int	Number of epoch. Required if train is true.
-	--data	one or more str	Dataset yaml configurations. If not given, will use predefined yaml in accordance with the version.
-	--weights	str	Model weight. Default yolov5s.pt.
1	--train	store_true	Train the model. If overwrite is true, it will remove the output (if exists) before training.
2	--test	store_true	Test the model. If overwrite is true, it will remove the test output (if exists) before testing.
3	--zip	store_true	Zip the output. If overwrite is true, it will remove the output zip (if exists) before zipping.
4	--upload	store_true	Upload the output zip.
5	--remove-zip	store_true	Remove the output zip, if exists.
6	--download	one or more url	Download the output zip. If overwrite is true, it will remove the output zip (if exists) before downloading.
7	--unzip	store_true	Unzip the output zip. If overwrite is true, it will remove the output (if exists) before unzipping.
8	--plot	store_true	Generate more plot. If train is true, this method will also be called.
9	--remove	store_true	Remove the output, if exists.

*Smaller priority number will be executed first

---

This project was developed as part of thesis project, Computer Science, BINUS University.