Downloading the Data

Before running the code, download the data from the Google Drive link and place it in the data directory, which should be located at the root of the project directory

Running the Code

To train the model, simply execute the training script:

python BoxHead.py

For evaluation, refer to the provided Jupyter Notebook rcnn.ipynb

Image plots of the proposals with the no-background classes and their ground truth boxes

The thick solid boxes represent the ground truth, while the thin dashed boxes are the proposals

Training and Validation curves of the total loss, the loss of the classifier, and the loss of the regressor of the Box Head

Please note that the plots start from epoch 0, so when referring to 'epoch 14' on the plot, it actually corresponds to epoch 15 in the training process

AP and mAP Scores of the Test Set

During the training process, we saved model checkpoints at epoch 15 and at the end of training, which is epoch 30. We obtained the following Average Precision (AP) values for the test set using these two models:

Epoch 15:

• The Average Precision for the 'Vehicles' class is: 0.923
• The Average Precision for the 'People' class is: 0.970
• The Average Precision for the 'Animals' class is: 0.937
• The mAP is 0.943

Epoch 30:

• The Average Precision for the 'Vehicles' class is: 0.931
• The Average Precision for the 'People' class is: 0.969
• The Average Precision for the 'Animals' class is: 0.936
• The mAP is 0.945

Image plots that contain the top 20 boxes produced by the Box Head before NMS

Image plots of the regressed boxes after the postprocessing

Analyzing mAP for a Single Batch of Data

This section provides an analysis of the mean Average Precision (mAP) for a single batch of data. The plot below illustrates the ground truth boxes in blue and the predicted boxes in red for the batch of images, offering insights into the model's performance on this specific subset of data.

The mAP of this batch of data is 0.852. Below, we present the precision-recall (PR) curves for each class, along with a detailed precision-recall table for the Vehicle Class

Precision-Recall Table for Vehicle Class

Image	Confidence	TP or FP	Precision	Recall
Image 6	0.99989796	TP	1/1	1/3
Image 3	0.99973518	TP	2/2	2/3
Image 3	0.99083507	TP	3/3	3/3
Image 3	0.99032766	FP	3/4	3/3
Image 4	0.98912036	FP	3/5	3/3

Note: The 'wiggles' in the precision-recall curve have been smoothed using interpolated precision, represented by the orange line, for the calculation of mAP, as illustrated in the figure below:

Implementation Specifics

• The threshold for NMS is set to 0.4, instead of the commonly used 0.5
• The optimizer is Adam optimizer with a learning rate of 0.0001
• A step learning rate scheduler reduces the learning rate by 10% every 10 epochs
• The effective batch size for training is set to 100, indicating the number of samples per mini-batch