Live demo
Run the model on your own image
The trained ResNet18 runs on a Hugging Face Space, embedded below. Upload a bird image to see the prediction, the Grad-CAM map, and the foreground / background attention-bias score.
Source:huggingface_space/app.py
What it reports
- Predicted class and confidence.
- Grad-CAM heatmap over the uploaded image.
- The 60% centre-crop foreground box.
- Foreground / background saliency split.
Images worth trying
- A landbird on grass, and a duck on water.
- A waterbird on a forest background.
- A landbird against strong blue water.
The conflict cases are where the background bias tends to surface.