Teach it where to look
V2 combines a penalty for evidence outside the object with support for evidence inside it. CNNs and Vision Transformers use their own class-evidence adapters.
01 / Attention learning
FCAL
Right answer. Wrong reason. Look at the bird.
A bird classifier can get the label right by looking at the scenery. FCAL asks a more interesting question: can we teach it to put its evidence on the bird itself?
Foreground-Constrained Attention Learning adds foreground supervision during training. The masks disappear at inference, leaving the same prediction interface and model size.
Research project / FCAL V2
Look beyond accuracy
The experiments compare clean predictions, class activation maps, and changed backgrounds. The bird benchmark includes three paired ResNet-50 training seeds.
Keep the trade-offs visible
Sharper localization does not guarantee higher accuracy. Results vary across datasets; the repository includes the ablations, baselines, and reproduction protocol.
Look at the evidence.
Straight from the project.
Adapted from the project README. Explore the repository for the full story.
Have something in mind?