Official implementation of "Finer-CAM".

CAM methods highlight image regions influencing predictions but often struggle in fine-grained tasks due to shared feature activation across similar classes. We propose Finer-CAM, which explicitly compares the target class with similar ones, suppressing shared features and emphasizing unique, discriminative details.

Finer-CAM retains CAM’s efficiency, offers precise localization, and adapts to multi-modal zero-shot models, accurately activating object parts or attributes. It enhances explainability in fine-grained tasks without increasing complexity.

Update

• 2026.04.19: Added a new Colab tutorial for using customized data: Finer-CAM Tutorial
• 2026.04.14: Added a new Hugging Face demo for the CUB classifier: Finer-CAM Demo
• 2025.03.13: Merged into jacobgil/pytorch-grad-cam, a wonderful library that supports multiple CAM-based methods.

Demo

Experience the power of Finer-CAM with our interactive demos! Witness accurate localization of discriminative features.

• Try the multi-modal demo and see how Finer-CAM activates detailed and relevant regions for diverse concepts:
  

• Test the CUB classifier demo to visualize fine-grained, discriminative traits with enhanced interpretability:
  

• Try the Colab tutorial and run Finer-CAM on your own data:
  

Requirements

Install the dependencies from this repo:

pip install -r requirements.txt

Run scripts and notebooks from the repository root so Python imports the local pytorch_grad_cam package in this tree.

Usage

Python API

FinerCAM is available from pytorch_grad_cam and wraps an existing CAM backend such as GradCAM. It keeps the normal CAM pipeline for collecting activations and gradients, but replaces the optimization target with the Finer-CAM objective.

from pytorch_grad_cam import FinerCAM, GradCAM
cam = FinerCAM(
    model=model,
    target_layers=target_layers,
    reshape_transform=reshape_transform,  # optional
    base_method=GradCAM,
)

Call FinerCAM with:

• input_tensor: input batch passed to the model.
• targets: optional list of pytorch_grad_cam target callables. See model_targets. If None, Finer-CAM targets are constructed automatically based on the model outputs.
• target_size: optional output size used when resizing CAM maps.
• eigen_smooth: enables eigenvalue-based smoothing in the wrapped CAM method.
• alpha: scaling factor used by FinerWeightedTarget for penalizing reference categories.
• reference_category_ranks: ranks from the similarity-sorted category list used to choose reference categories when targets=None. The default [1, 2, 3] uses the second to fourth most similar categories as references. If a requested rank exceeds the number of available classes, it is ignored.
• target_idx: the index of the target category, usually the ground-truth category. If omitted, the highest-scoring category in each sample is used.
• H, W: optional feature-grid height and width for backbones that need them in the activation/gradient path, such as ViT-style reshape transforms.

FinerCAM.forward(...) returns a tuple:

cam_map, outputs, main_categories, references = cam(
    input_tensor=input_tensor,
    targets=None,
    alpha=1.0,
    reference_category_ranks=[1, 2, 3],
    target_idx=target_idx,
    H=grid_height,
    W=grid_width,
)

• cam_map: aggregated CAM map from the wrapped backend.
• outputs: raw model outputs from the forward pass.
• main_categories: automatically selected main category per sample when targets=None.
• references: automatically selected reference categories per sample when targets=None.

When targets=None, Finer-CAM computes a similarity ranking by sorting class logits according to their absolute distance from the reference logit. The closest class becomes the main category and the selected reference_category_ranks become the reference set.

FinerWeightedTarget

Automatic target construction uses FinerWeightedTarget, which implements the weighted relative objective used by Finer-CAM. For a main category n and a reference set i, it computes