Semi-Supervised Pseudo Labeling with Object Detection Models

• Images Capturing Pipeline

  • Images Capture with Augmentaion image_capture/capture_images.py

    cap = Image_Capture(0) #camera device id
    
    '''
    args : 
            img_size : Capture image size (i.e 608x608)
            img_dir : Image Folder
            rotate90 : rotate image by 90 degree angle
            rotate180 : rotate image by 180 degree angle
            rotate270 : rotate image by 270 degree angle
            scale : Scale the image
            scale_val : Scale value of the image
    '''
    cap.capture_image(img_size, 
                      img_dir,
                      rotate90=True,rotate180=True,rotate270=True,
                      scale=True,scale_val=0.2)
    

• Data Augmentaion Pipeline for Object Detection

  • Data Augmentation For Object Detection

    git clone git@github.com:LahiRumesh/Object-Detection_Data-Augmentation.git
    cd Object-Detection_Data-Augmentation/

Pytorch YOLO models Train

• Data Prepare

  • annotation file should be in vott csv format

  image	xmin	ymin	xmax	ymax	label
  image1.jpg	50	150	288	328	label1
  image1.jpg	300	263	410	333	label2
  image2.jpg	88	63	110	223	label1
  image3.jpg	22	190	150	250	label3

  • Data Folder - >

    • image1.jpg image2.jpg image3.jpg annotation.csv

• Use the train_models.py script to train YOLOv3 and YOLOv4 models

• Training arguments for the model training.

  • --model : object detection model, i.e. YOLOv4 or YOLOv3

  • --data_dir : Image folder which contains images and the csv file

  • --weights : pre-trained weights path

  • --validation : validation data split

  • --epochs : number of training epochs

  • --batch_size : train and validation batch size

  • --image_size : train and test image size (should be multiply by 32 i.e (416,416),(512,512) or (608,608) )

  • --learning_rate: learning rate

  • --device : cuda device, i.e. 0 or 0,1,2,3 or cpu

Pytorch model Inference with Pseudo Labeling

• Use the pseudo_label.py script to for the pseudo labeling

• Inference arguments for the pseudo labeling.

  • --checkpoint : saved checkpoint weight file path

  • --class_file : class file which contains class names i.e class.names

  • --dir_name : folder path which contians unlabeld images

  • --conf_thresh : confident threshold value

  • --iou_thresh : IOU threshold value

  • --image_size : test image size (should be multiply by 32 i.e (416,416),(512,512) or (608,608) )

Reference:

• https://github.com/ayooshkathuria/pytorch-yolo-v3
• https://github.com/eriklindernoren/PyTorch-YOLOv3
• https://github.com/Tianxiaomo/pytorch-YOLOv4

  @article{yolov3,
  title={YOLOv3: An Incremental Improvement},
  author={Redmon, Joseph and Farhadi, Ali},
  journal = {arXiv},
  year={2018}
  }

@article{yolov4, title={YOLOv4: YOLOv4: Optimal Speed and Accuracy of Object Detection}, author={Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao}, journal = {arXiv}, year={2020} }