| Backbone |
ResNet-50 with FPN (Feature Pyramid Network) |
| Pretrained Weights |
Trained on ImageNet for feature extraction. |
| RPN (Region Proposal Network) |
Generates region proposals based on extracted features from the backbone. |
| ROI Align |
Aligns region proposals to a fixed size for consistent feature extraction. |
| Box Head |
Fully connected layers for refining bounding boxes and classifying objects. |
| Box Predictor |
Replaced with a custom predictor: FastRCNNPredictor for handling custom classes. |
| Number of Classes |
Configurable (including background). |
| Loss Function |
Combines classification and regression losses for multi-task optimization. |
| Optimizer |
Stochastic Gradient Descent (SGD) with momentum for optimization. |
| Learning Rate Scheduler |
StepLR to decay learning rate every few epochs for better convergence. |
| Batch Normalization |
Applied within the backbone for stable training. |
| Data Format |
Input: Tensor of shape (Batch Size, Channels, Height, Width) in PyTorch's NCHW format. |
| Output |
- Class probabilities for each region proposal. |
|
- Refined bounding box coordinates for each detected object. |
