Pretraining YOLOv5 Backbone: Cross Stage Partial Convolutions DarkNet with Spatial Pyramid Pooling

• Self supervised autoencoder approach aimed towards improving the detections of SOTA object detector YOLOv5 by detecting and re-constructing multiple objects in an image.
• This approach effectively performs semantic segmentation without pixel-wise annotations but uses only the bounding box coordinates.
• Apart from which, the project aims at contributing to Ultralytics YOLOv5 by constructing a method of transferring only backbone weights to models for training and fine-tuning.

Methdology

• We pre-train the backbone of YOLOv5 from scratch and by transfering the weights of yolov5s.pt.
• The loss function is an L2 Pixel-Wise Loss with additional focus on pixels belonging to a bounding box
• The keys of the backbone of this model is then matched with the Ultralytics YOLOv5 state dictionary structure enabling us to transfer these weights to Ultralytics YOLOv5 and get bounding box coordinates for all the objects in the image.
• The Global Wheat 2021 dataset was used for all experiments

Architecture

Architecture	Description
Type	Autoencoder
Backbone	CSP Darknet with Spatial Pyramid Pooling
Decoder	Transpose CSP Darknet

Directory Structure

├── Outputs
│   ├── *.png
├── Helpers
│   ├── mask.py
│   ├── blur.py
│   ├── match_keys.py
├── Scripts
│   ├── Model
│   ├──   ├── model_segments
│   ├──   ├── │   ├──modules.py
│   ├──   ├── backbone.py
│   ├──   ├── combined.py
│   ├──   ├── decoder.py
│   ├── dataset.py
│   ├── initialize_weights.py
│   ├── keys.py
│   ├── test.py
│   ├── train.py
├── Scripts_COCO
│   ├── Model
│   ├──   ├── model_segments
│   ├──   ├── │   ├──modules.py
│   ├──   ├── backbone.py
│   ├──   ├── combined.py
│   ├──   ├── decoder.py
│   ├── dataset.py
│   ├── initialize_weights.py
│   ├── keys.py
│   ├── test.py
│   ├── train.py
├── requirements.txt
├── readme.md
└── .gitignore

Training and Testing

• Clone the repository
• Create a .env file with variables as described below and create the necessary directories.
• You may run train and test scripts for both with and without pretrained weights
• Run mask.py to generate the mask for each image
• Create directories to save checkpoints and edit the paths as required in the train.py files.
• To train from Scratch, run the train.py file under Scripts, for fine-tuning yolov5s backbone, run train.py under Scripts_COCO
• After generating the weights, run the keys.py file to match the state dictionary keys to that of YOLOv5.

Environment Variables

Variable	Description
TRAIN_Y_PATH	Path to Train Dataset of Global Wheat Head,2021
MASK	Path to generated mask based on bounding box coordinates