Consensus Focus for Object Detection and minority classes

TL;DR

This work tackles semi-supervised and long-tailed object detection by addressing negative transfer through a modified consensus focus that uses a knowledge-vote style voting scheme to weight detectors by confidence. The approach combines per-source detections with Weighted Box Fusion, augments the source pool with extended consensus data $D^{(I+1)}$, and computes source contributions via a consensus quality metric $Q(S')$ and marginal contributions $CF(D^{(i)})$ to derive reweighting factors $\alpha_i^{CF}$. Empirical evaluation on synthetic driving datasets shows higher confidence and more accurate bounding boxes than standard NMS, soft-NMS, and WBF, especially at practical thresholds, albeit with ~3× slower runtime due to the combinatorial consensus analysis. The method offers robust performance in low-label regimes and provides a pathway toward federated learning and semi-supervised extensions in multi-domain detection tasks.

Abstract

Ensemble methods exploit the availability of a given number of classifiers or detectors trained in single or multiple source domains and tasks to address machine learning problems such as domain adaptation or multi-source transfer learning. Existing research measures the domain distance between the sources and the target dataset, trains multiple networks on the same data with different samples per class, or combines predictions from models trained under varied hyperparameters and settings. Their solutions enhanced the performance on small or tail categories but hurt the rest. To this end, we propose a modified consensus focus for semi-supervised and long-tailed object detection. We introduce a voting system based on source confidence that spots the contribution of each model in a consensus, lets the user choose the relevance of each class in the target label space so that it relaxes minority bounding boxes suppression, and combines multiple models' results without discarding the poisonous networks. Our tests on synthetic driving datasets retrieved higher confidence and more accurate bounding boxes than the NMS, soft-NMS, and WBF. The code used to generate the results is available in our GitHub repository: http://github.com/ErikValle/Consensus-focus-for-object-detection.

Figures (2)

• Figure 1: Knowledge vote ensemble for object detection.
• Figure 2: F1 curves comparing the performance of our method versus NMS, Soft-NMS, and WBF (all weights equal to one) after skipping boxes with confidence lower than 0.0001.