Class Activation Mapping (CAM) - Discriminative Localization

This project demonstrates Class Activation Mapping (CAM) to enable discriminative localization — letting a CNN not only classify an image but also highlight the regions responsible for the decision.

---

Traditional CNNs give you what the class is, but not where or why. CAM bridges this gap between prediction and interpretation.

---

1. Discriminative Localization

• Goal: Highlight which part of the image led to the prediction.
• Traditional CNNs output class labels but give no insight into the spatial regions influencing the decision.
• Example: A CNN predicts "cat", but CAM shows the cat’s face contributed most.

---

2. Problem Addressed

• CNNs are black boxes — you get a prediction, but no explanation.
• CAM provides interpretability by visualizing regions important for classification.
• Example: In medical imaging, showing “where” the tumor is builds trust.

---

3. Class Activation Mapping (CAM)

• Localizes class-specific regions using feature maps and classifier weights.

• Equation (intuition):

  [ \text{CAM}_c(x, y) = \sum_k w_k^c f_k(x, y) ]

  • ( w_k^c ): Weight for class ( c )
  • ( f_k(x, y) ): Activation at location ( (x, y) ) in feature map ( k )

• Example: CAM for "golden retriever" highlights the head and ears.

---

4. Global Average Pooling (GAP)

• Replaces fully connected (FC) layers to preserve spatial structure.
• GAP takes the average of each feature map → outputs a vector → passed to a linear classifier.
• Example: From feature map (512, 7, 7) → GAP → (512,) → Linear → Softmax
• Benefits:
  • Retains spatial info
  • Reduces overfitting
  • Enables CAM computation

---

5. Weakly Supervised Localization

• No need for bounding boxes — only image-level labels.
• The model learns to localize during classification training.
• Example: Labeled "dog" → after training, CAM shows where the dog is.

---

6. Key Structural Changes

Component	Traditional CNN	CAM-based CNN
FC Layers	Present (e.g., 4096)	Removed
GAP Layer	Absent	Inserted before classifier
Classifier	Multi-layer FC	Single Linear layer
Localization	Not possible	Enabled via CAM

---

Outcomes

Object Localization

• CAM allows drawing bounding boxes around the most activated region.
• Example: A bird’s head can be localized without box annotations.

Fine-Grained Recognition

• Differentiates visually similar classes using discriminative parts.
• Example: Distinguish "Pug" vs. "Bulldog" using facial features.

---

🛠 Implementation Notes

• Dataset Used: Oxford-IIIT Pet Dataset

  • Rich in fine-grained categories.
  • CIFAR-10 (32x32) was too low-res for effective CAM visualization.

• Model Setup:

  • Forward hook added after the 4th convolutional block (just before GAP).
  • Feature maps are saved during the forward pass.
  • Class weights from the final linear layer are used to compute CAM.

• CAM Generation:

  • Weighted sum of feature maps using the predicted class’s weights.
  • Normalized and upsampled to original image size for visualization.

---