Predicts bounding box coordinates directly without using box generator. By directly predicting bounding boxes, accurate box coordinates can be obtained.

        for reg_layer, conf_layer in zip(self.regression_headers, self.classification_headers):
            locs = reg_layer(locs)  # (B, 4, H, W)
            confs = conf_layer(confs)  # (B, num_classes, H, W)

The PostProcessor class handles post-processing for the bounding boxes and scores predicted by the model.

• Inputs:

  • locs: List of bounding box coordinates
  • confs: List of class confidence scores

• Outputs:

  • result_boxes: Final selected bounding boxes
  • result_labels: Class labels of the bounding boxes
  • result_scores: Confidence scores of the bounding boxes
  • result_confs: Raw confidence values

• Functionality:

  • Applies softmax to select the highest-scoring class
  • Removes low-score predictions based on score_threshold
  • Uses Non-Maximum Suppression (NMS) to eliminate redundant boxes

The SSD class consists of the following major components:

• Backbone: CNN network for feature map extraction

• Extras: Additional layers to detect objects of different sizes

• Classification Headers: Predict class scores

• Regression Headers: Predict bounding box coordinates

• PostProcessor: Filters and processes detection results

• Inputs:

  • x: Input image tensor

• Outputs:

  • result_boxes: Bounding box coordinates
  • result_labels: Class labels
  • result_scores: Scores of detected objects
  • result_confs: Raw confidence values

The loss function used for training the SSD model.

• Loss Function Components:

  • SmoothL1Loss: Bounding box regression loss
  • CrossEntropyLoss: Classification loss

• Inputs:

  • result_boxes: Predicted bounding boxes
  • result_confs: Predicted class confidence scores
  • target_locs: Ground truth bounding boxes
  • target_labels: Ground truth class labels

• Outputs:

  • total_loss: Total loss (regression loss + classification loss)
  • train_acc: Training accuracy

• Hungarian Matching for Alignment:

  • Matches predicted and ground truth bounding boxes before computing loss

python train_ssd.py --arch mobilenetv3_ssd --datapath dataset_coco --epochs 10 --batch-size 5 --lr 0.001 --save ssd.pt