🧠 DETR Interpretability Dashboard with Controls

How to use this dashboard

• Upload an image using the left panel. The model will run object detection and list detected objects. Try imageNet
• Use the "Confidence Threshold" slider to filter detections by score. Detections below the threshold are hidden.
• Pick a detection from the dropdown to generate explanations for that object.
• Choose between Grad-CAM and Grad-CAM++ (Grad-CAM++ often gives sharper, more localized maps).
• MC Dropout Samples controls how many stochastic forward passes are used to estimate prediction uncertainty. More samples give smoother estimates but take longer.
• Dropout Probability sets the dropout rate used during MC Dropout; higher values typically increase predicted uncertainty.

Tooltips are provided on each control (hover or focus) for quick hints.

Control tooltips (quick reference)

• Confidence Threshold: Filter out detections with confidence below this value.
• Grad-CAM Variant: Choose the gradient-based visualization method. Grad-CAM++ may highlight smaller regions more precisely.
• MC Dropout Samples: Number of stochastic forward passes for uncertainty estimation. Increase for more stable results.
• Dropout Probability: Dropout rate used during MC Dropout sampling. Higher values typically increase predictive variance.
• Pick a detection: Select which detected object to explain. Format shown as 'index: label (score)'.

Key interpretability choices & why they matter

• Baseline (Integrated Gradients): Defines what 'no signal' looks like. Black (zeros) is simple, but blurred or neutral baselines may give more meaningful attributions.
• Which conv layer for Grad-CAM: Early layers give fine texture but low semantics; very late layers are coarse. A late backbone conv (default used) is a good compromise.
• Number of MC Dropout samples: More samples = smoother, more stable uncertainty estimates, but higher compute cost.
• Grad-CAM vs Grad-CAM++: Grad-CAM++ can be sharper and better for overlapping instances; vanilla Grad-CAM is faster and simpler.

Further reading (recommended)

• Grad-CAM — Selvaraju et al., 2017 (arXiv) — the original Grad-CAM paper; explains the core idea of gradient-weighted localization.
• Grad-CAM++ — Chattopadhay et al. — an improved variant that often produces sharper maps and handles multiple instances better.
• Visualizing the Impact of Feature Attribution Baselines (Distill) — an accessible deep dive on baseline choices for Integrated Gradients.
• Captum docs — IntegratedGradients — practical API notes for baseline, n_steps, and convergence delta.
• Constructing sensible baselines for Integrated Gradients — discussion and techniques for choosing baselines beyond a black image.
• A New Baseline Assumption of Integrated Gradients Based on Shapley Values — recent research on improved baselines.