Faster-RCNN finetuning with PyTorch. Object detection using PyTorch. Custom dataset. Wheat detection

In this video, we are going to see how can we fine tune a pretrained faster-rcnn model using PyTorch. The dataset we will be using is the wheat detection dataset.

The video has three objectives:
1. Explain the main portion of the code for fine tuning faster-rcnn.
2. How does the faster-rcnn finetuning look like and what layers are affected as a result?
3. See the whole code for fine tuning faster-rcnn.

We will see how the last layer of the faster-rcnn module is affected due to fine tuning. We also see how the Non Max Suppression (NMS) also works in the context of fine tuning faster-rcnn.

The main code for faster-rcnn is:

model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained = True)
num_classes = 2
in_features = model.roi_heads.box_predictor.cls_score.in_features
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)