Stop Sign Detection with Fine-tuned RCNN (MATLAB)

This repository contains MATLAB code for fine tuning a pre-trained RCNN model. Below is a breakdown of the functionalities provided:

Training Script:

Load Data:

• load gTruth.mat: Loads ground truth data containing bounding box information and labels for stop signs in your training dataset.
• load('rcnnStopSigns.mat', 'layers'): Loads a pre-trained RCNN model architecture stored in the rcnnStopSigns.mat file.

Define Training Options:

• trainingOptions('sgdm', ...): Sets the training optimizer to stochastic gradient descent (sgdm).
• Other options define hyperparameters for training, including:
  • 'MiniBatchSize', 32: Batch size of images used for training updates.
  • 'InitialLearnRate', 1e-6: Initial learning rate for adjusting model weights.
  • 'MaxEpochs', 2: Maximum number of times to iterate through the entire training dataset.
  • 'ExecutionEnvironment', 'gpu': Leverages GPU for faster training (if available).

Train RCNN Model:

• trainRCNNObjectDetector(gTruth, layers, Options, 'NegativeOverlapRange', [0 0.1]): Trains the RCNN object detector.
  • gTruth: Ground truth data for stop signs.
  • layers: Pre-trained RCNN model architecture.
  • Options: Training options defined earlier.
  • 'NegativeOverlapRange', [0 0.1]: Defines the range of bounding box overlaps considered negative examples during training (boxes with less than 10% overlap with a ground truth stop sign are considered negative).

Testing Script:

Load Test Image:

• uigetfile: Prompts the user to select a test image.
• imread: Reads the selected image into the workspace.

Stop Sign Detection:

• detect(rcnn, img, 'MiniBatchSize', 32): Uses the trained RCNN model (rcnn) to detect stop signs in the test image (img). It also specifies a mini-batch size of 32 for efficient processing during detection.
  • The output includes:
    • bbox: Bounding boxes around detected stop signs.
    • score: Confidence scores for each detection (probability of being a stop sign).
    • label: Labels for each detection (should be 'stop sign' in this case).

Filter and Visualize Results:

• nobox=size(score, 1): Counts the number of detected objects.
• scorebox=[score, bbox]: Combines confidence scores with bounding box information.
• scorebox = sortrows(scorebox, 'descend'): Sorts detections by confidence score (highest score first).
• The code iterates through detections and displays only those with a confidence score above a threshold (e.g., 0.95). It then visualizes the test image with bounding boxes and confidence scores for detected stop signs.

Additional Notes:

• This is a basic implementation. You can modify it to:
  • Train the model for a longer duration or with different hyperparameters for better accuracy.
  • Evaluate the model's performance on a separate validation dataset.
  • Integrate this code into a larger application for real-time stop sign detection in video streams.