YOLOv7 Grad-CAM and CAM Visualizations

This repository provides a simple way to visualize where a YOLOv7 model focuses when making predictions, using Class Activation Mapping (CAM) techniques such as Grad-CAM, Grad-CAM++, and XGrad-CAM. These visualizations help explain model decisions by highlighting important regions in an image.

What are CAM and Grad-CAM? (Brief)

• Class Activation Maps (CAMs): Visual explanations that show which image regions contribute most to a model’s decision for a target class. CAMs are useful for trust, transparency, and debugging, especially in high-stakes domains (e.g., medical).
• How CNNs enable CAMs: Early CNN layers learn simple features (edges, lines), while deeper layers learn higher-level patterns. CAM-based methods leverage the last convolutional feature maps to produce localization heatmaps that explain predictions.
• Grad-CAM (Gradient-weighted CAM): Computes gradients of a target class score with respect to activations in a chosen convolutional layer, aggregates them into neuron-importance weights, and forms a heatmap that highlights influential regions. Grad-CAM variants (e.g., Grad-CAM++, XGrad-CAM, Layer-CAM) can improve localization sharpness or flexibility across layers.

Examples

• Grad-CAM example heatmap overlay:

• Occlusion technique illustration: masking different regions to see how the prediction changes (red = important, blue = less important):

Requirements

• Python 3.x
• PyTorch
• OpenCV
• NumPy
• Matplotlib
• tqdm
• PIL (Pillow)
• pytorch_grad_cam

Usage

1. Prepare YOLOv7 assets

   • Download the YOLOv7 weights file (e.g., yolov7.pt) and ensure the configuration file exists (e.g., cfg/training/yolov7.yaml).

2. Adjust parameters if needed

   • Edit get_params inside gradcam.py to change defaults such as weight, cfg, device, CAM method, target layer, backward type, confidence threshold, and ratio.

3. Run the script

   • The minimal command will use the defaults and process the example image:

   python gradcam.py

   • By default, it reads inference/images/image3.jpg and writes outputs into the result directory.

Parameters

• weight: Path to the YOLOv7 weights file.
• cfg: Path to the YOLOv7 configuration file.
• device: Compute device ('cpu' or 'cuda').
• method: CAM method ('GradCAM', 'GradCAMPlusPlus', 'XGradCAM').
• layer: Target layer to visualize (e.g., 'model.model[-2]').
• backward_type: Backward objective ('class' for class score or 'conf' for confidence).
• conf_threshold: Confidence threshold for detections.
• ratio: Fraction of top predictions to visualize.

Example (inside gradcam.py)

def get_params():
    params = {
        'weight': 'yolov7.pt',
        'cfg': 'cfg/training/yolov7.yaml',
        'device': 'cpu',
        'method': 'GradCAM',  # GradCAMPlusPlus, GradCAM, XGradCAM
        'layer': 'model.model[-2]',
        'backward_type': 'class',  # class or conf
        'conf_threshold': 0.6,  # 0.6
        'ratio': 0.02  # 0.02-0.1
    }
    return params

if __name__ == '__main__':
    model = yolov7_heatmap(**get_params())
    model('inference/images/image3.jpg', 'result')

Notes

• You can choose different layers for visualization; deeper layers often give more semantic heatmaps, while earlier layers may show lower-level features.
• Grad-CAM and its variants are widely used to understand, debug, and increase trust in CNN-based models by revealing which regions drive predictions.