Skip to content

Latest commit

 

History

History
93 lines (71 loc) · 8.98 KB

README.md

File metadata and controls

93 lines (71 loc) · 8.98 KB
Raw

Image Augmentation for Object Detection and Segmentation with Diffusion Models

Authors: Leon Useinov, Valeria Efimova, Sergey Muravyov

image

We propose an image augmentation approach for object detection and segmentation tasks based on the Stable Diffusion XL diffusion model. This repository contains code to verify our research. The following things are implemented:

  • Preprocessing scripts for PCB Defects and Potholes datasets.
  • Inpainting (augmentation) scripts for MVTec AD Bottle, PCB Defects and Potholes datasets.
  • Postprocessing scripts for MVTec AD Bottle, PCB Defects and Potholes datasets.
  • YOLOv8n training scripts.

Requirements

Docker and Docker Compose.

Installation

git clone https://github.com/PnthrLeo/diffusion-augmentation.git
cd diffusion-augmentation
docker compose build

Models

Download:

  1. SDXL Base (put it in models/comfyUI_models/checkpoints/sd_xl_base_1.0.safetensors)
  2. SDXL Refiner (put it in models/comfyUI_models/checkpoints/sd_xl_refiner_1.0.safetensors)
  3. ControlNet Depth for SDXL (put it in models/comfyUI_models/controlnet/controlnet-depth-sdxl-1.0.safetensors)
  4. ControlNet Canny for SDXL (put it in models/comfyUI_models/controlnet/controlnet-canny-sdxl-1.0.safetensors)
  5. CLIP Vision (put it in models/comfyUI_models/clip_vision/CLIP-ViT-H-14.safetensors)
  6. IP-Adapter (put it in models/IPAdapter_models/ip-adapter-plus_sdxl_vit-h.safetensors)

Datasets

Download:

  1. MVTec AD dataset from Kaggle (put archive contents in data/mvtec/orig/)
  2. PCB Defects dataset from Roboflow in YOLOv8 format (put archive contents in data/pcb/orig)
  3. Potholes dataset from Roboflow in YOLOv8 format (put archive contents in data/pothole/orig)

Usage

First, run preprocessing scripts (only for PCB Defects and Potholes datasets):

# Example of running preprocessing for PCB Defects dataset
ORIG_DATA_PATH=data/pcb/orig PREPROCESSING_SCRIPT=pcb.py  docker compose -f docker-compose-preprocessing.yaml up

# Example of running preprocessing for Potholes dataset
ORIG_DATA_PATH=data/pothole/orig PREPROCESSING_SCRIPT=pothole.py  docker compose -f docker-compose-preprocessing.yaml up

Second, run inpainting scripts (ATTENTION: CONTAINER SHOULD BE STOPPED MANUALLY AFTER INPAINTING COMPLETION):

# Example of running inpainting (1, 2, 3 versions) for MVTec AD Bottle dataset
DATASET_PATH=data/mvtec/orig INPAINTING_OUTPUT_PATH=data/mvtec/mvtec-inpainting/mvtec-inpainting-1 INP_IMG_PER_ORIG_IMG=15 INPAINTING_SCRIPT=mvtec_perform_inpainting_1_0.py  docker compose up
DATASET_PATH=data/mvtec/orig INPAINTING_OUTPUT_PATH=data/mvtec/mvtec-inpainting/mvtec-inpainting-2 INP_IMG_PER_ORIG_IMG=15 INPAINTING_SCRIPT=mvtec_perform_inpainting_2_0.py  docker compose up
DATASET_PATH=data/mvtec/orig INPAINTING_OUTPUT_PATH=data/mvtec/mvtec-inpainting/mvtec-inpainting-3 INP_IMG_PER_ORIG_IMG=15 INPAINTING_SCRIPT=mvtec_perform_inpainting_3_0.py  docker compose up

# Example of running inpainting (1, 2, 3 versions) for PCB Defects dataset
DATASET_PATH=data/pcb/orig INPAINTING_OUTPUT_PATH=data/pcb/pcb-inpainting/pcb-inpainting-1 INP_IMG_PER_ORIG_IMG=6 INPAINTING_SCRIPT=pcb_perform_inpainting_1_0.py  docker compose up
DATASET_PATH=data/pcb/orig INPAINTING_OUTPUT_PATH=data/pcb/pcb-inpainting/pcb-inpainting-2 INP_IMG_PER_ORIG_IMG=6 INPAINTING_SCRIPT=pcb_perform_inpainting_2_0.py  docker compose up
DATASET_PATH=data/pcb/orig INPAINTING_OUTPUT_PATH=data/pcb/pcb-inpainting/pcb-inpainting-3 INP_IMG_PER_ORIG_IMG=6 INPAINTING_SCRIPT=pcb_perform_inpainting_3_0.py  docker compose up

# Example of running inpainting (1, 2, 3 versions) for Potholes dataset
DATASET_PATH=data/pothole/orig INPAINTING_OUTPUT_PATH=data/pothole/pothole-inpainting/pothole-inpainting-1 INP_IMG_PER_ORIG_IMG=6 INPAINTING_SCRIPT=pothole_perform_inpainting_1_0.py  docker compose up
DATASET_PATH=data/pothole/orig INPAINTING_OUTPUT_PATH=data/pothole/pothole-inpainting/pothole-inpainting-2 INP_IMG_PER_ORIG_IMG=6 INPAINTING_SCRIPT=pothole_perform_inpainting_2_0.py  docker compose up
DATASET_PATH=data/pothole/orig INPAINTING_OUTPUT_PATH=data/pothole/pothole-inpainting/pothole-inpainting-3 INP_IMG_PER_ORIG_IMG=6 INPAINTING_SCRIPT=pothole_perform_inpainting_3_0.py  docker compose up

Third, run postprocessing scripts (to get final training datasets):

# Example of running postprocessing for MVTec AD Bottle dataset (1, 2, 3 versions)
ORIG_DATA_PATH=data/mvtec/orig INPAINTED_DATA_PATH=data/mvtec/mvtec-inpainting/mvtec-inpainting-1 FINAL_DATASET_PATH=data/mvtec/mvtec-datasets/mvtec-with-inpainting-1 INP_IMG_PER_ORIG_IMG=15 NOT_INPAINTED_DATA_PATH=data/mvtec/mvtec-datasets/mvtec-no-inpainting POSTPROCESSING_SCRIPT=mvtec.py  docker compose -f docker-compose-postprocessing.yaml up
ORIG_DATA_PATH=data/mvtec/orig INPAINTED_DATA_PATH=data/mvtec/mvtec-inpainting/mvtec-inpainting-2 FINAL_DATASET_PATH=data/mvtec/mvtec-datasets/mvtec-with-inpainting-2 INP_IMG_PER_ORIG_IMG=15 NOT_INPAINTED_DATA_PATH=data/mvtec/mvtec-datasets/mvtec-no-inpainting POSTPROCESSING_SCRIPT=mvtec.py  docker compose -f docker-compose-postprocessing.yaml up
ORIG_DATA_PATH=data/mvtec/orig INPAINTED_DATA_PATH=data/mvtec/mvtec-inpainting/mvtec-inpainting-3 FINAL_DATASET_PATH=data/mvtec/mvtec-datasets/mvtec-with-inpainting-3 INP_IMG_PER_ORIG_IMG=15 NOT_INPAINTED_DATA_PATH=data/mvtec/mvtec-datasets/mvtec-no-inpainting POSTPROCESSING_SCRIPT=mvtec.py  docker compose -f docker-compose-postprocessing.yaml up

# Example of running postprocessing for PCB Defects dataset (1, 2, 3 versions)
ORIG_DATA_PATH=data/pcb/orig INPAINTED_DATA_PATH=data/pcb/pcb-inpainting/pcb-inpainting-1 FINAL_DATASET_PATH=data/pcb/pcb-datasets/pcb-with-inpainting-1 INP_IMG_PER_ORIG_IMG=6 NOT_INPAINTED_DATA_PATH=data/pcb/pcb-datasets/pcb-no-inpainting POSTPROCESSING_SCRIPT=pcb.py  docker compose -f docker-compose-postprocessing.yaml up
ORIG_DATA_PATH=data/pcb/orig INPAINTED_DATA_PATH=data/pcb/pcb-inpainting/pcb-inpainting-2 FINAL_DATASET_PATH=data/pcb/pcb-datasets/pcb-with-inpainting-2 INP_IMG_PER_ORIG_IMG=6 NOT_INPAINTED_DATA_PATH=data/pcb/pcb-datasets/pcb-no-inpainting POSTPROCESSING_SCRIPT=pcb.py  docker compose -f docker-compose-postprocessing.yaml up
ORIG_DATA_PATH=data/pcb/orig INPAINTED_DATA_PATH=data/pcb/pcb-inpainting/pcb-inpainting-3 FINAL_DATASET_PATH=data/pcb/pcb-datasets/pcb-with-inpainting-3 INP_IMG_PER_ORIG_IMG=6 NOT_INPAINTED_DATA_PATH=data/pcb/pcb-datasets/pcb-no-inpainting POSTPROCESSING_SCRIPT=pcb.py  docker compose -f docker-compose-postprocessing.yaml up

# Example of running postprocessing for Potholes dataset (1, 2, 3 versions)
ORIG_DATA_PATH=data/pothole/orig INPAINTED_DATA_PATH=data/pothole/pothole-inpainting/pothole-inpainting-1 FINAL_DATASET_PATH=data/pothole/pothole-datasets/pothole-with-inpainting-1 INP_IMG_PER_ORIG_IMG=6 NOT_INPAINTED_DATA_PATH=data/pothole/pothole-datasets/pothole-no-inpainting POSTPROCESSING_SCRIPT=pothole.py  docker compose -f docker-compose-postprocessing.yaml up
ORIG_DATA_PATH=data/pothole/orig INPAINTED_DATA_PATH=data/pothole/pothole-inpainting/pothole-inpainting-2 FINAL_DATASET_PATH=data/pothole/pothole-datasets/pothole-with-inpainting-2 INP_IMG_PER_ORIG_IMG=6 NOT_INPAINTED_DATA_PATH=data/pothole/pothole-datasets/pothole-no-inpainting POSTPROCESSING_SCRIPT=pothole.py  docker compose -f docker-compose-postprocessing.yaml up
ORIG_DATA_PATH=data/pothole/orig INPAINTED_DATA_PATH=data/pothole/pothole-inpainting/pothole-inpainting-3 FINAL_DATASET_PATH=data/pothole/pothole-datasets/pothole-with-inpainting-3 INP_IMG_PER_ORIG_IMG=6 NOT_INPAINTED_DATA_PATH=data/pothole/pothole-datasets/pothole-no-inpainting POSTPROCESSING_SCRIPT=pothole.py  docker compose -f docker-compose-postprocessing.yaml up

Acknowledgements

The research was supported by the ITMO University, project 623097 ”Development of libraries containing perspective machine learning methods”.