Software Implementation for: Neural-network–enabled passive localization of multiple gamma emitters in radioactive waste

Bugfixes Evaluation: false positive retention in matching Previously, the prediction matching loop terminated after filling all ground-truth slots, silently discarding any remaining predicted boxes. These unmatched predictions are now retained and placed into empty slots, so they correctly contribute to union volume and confidence metrics. This fix yields a sound false-positive rate and, in particular, corrects Total Volume IoU, which was artificially inflated in earlier versions. Evaluation: consistent cost metric between training and evaluation Matching during evaluation previously used L2 (Euclidean) distance. It now uses the same hybrid Smooth L1 + CIoU cost as the training loss, so that the assignment used to compute R² and IoU metrics is consistent with the assignment the model was trained on. Evaluation: float tolerance for padded slot detection. Padded slots (value = −1) were previously identified by exact floating-point equality, which could fail due to precision artefacts. All masking in evaluation now uses np.isclose(..., atol=1e-3). Improvements Confidence-weighted bipartite matching in the loss The assignment cost matrix now incorporates the model's predicted confidence as a penalty (match_cost = cost_pure − scale × conf_penalty), following standard practice in set-prediction architectures (DETR, YOLOv10). When multiple predicted boxes are spatially equivalent, the one with the highest confidence is stably assigned the positive target, resolving confidence score collapse during training. Confidence target assignment via matching indices Confidence targets are now derived from the same bipartite matching indices as the box targets, guaranteeing that each prediction slot receives the correct binary confidence label. Previously, confidence targets were assigned independently of the box matching.