FreeAnchor: Learning to Match Anchors for Visual Object Detection

Xiaosong Zhang, Fang Wan, Chang Liu, Rongrong Ji, Qixiang Ye

Abstract:

Modern CNN-based object detectors assign anchors for ground-truth objects under the restriction of object-anchor
Intersection-over-Unit (IoU). In paper FreeAnchor [1], it proposed a learning-to-match approach and breaks the IoU
restriction, which allows objects to match anchors in a flexible manner. For a better understanding of Freeanchor,
we conducted different experiments based on the code published by FreeAnchor. First, we reproduced the baseline
result and found this network is robust according to consistent results. Second, we did two ablation experiments by
changing two components in this network, which are saturated linear function and mean-max function respectively.
Basically, FreeAnchor updates hand-crafted anchor assignment to “free” anchor matching by formulating detector
training as a maximum likelihood estimation (MLE) procedure. It targets learning features which best explain a
class of objects in terms of both classification and localization [1]. FreeAnchor proposed a detection customized
likelihood including precision and recall and improved them with a likelihood optimization process. In this paper,
we conducted different experiments based on this FreeAnchor method to demonstrate its reproducibility. Overall,
we reached the baseline published in the paper, and performed ablation experiments and hyperparameters tuning
which showed the network’s robustness.

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=Bkf97NrlIH