Title

CORRELATION LOSS: ENFORCING CORRELATION BETWEEN CLASSIFICATION AND LOCALIZATION IN OBJECT DETECTION

Abstract

Abstract

Object detectors are conventionally trained by a weighted sum of classification and localization losses. Recent studies (e.g., predicting IoU with an auxiliary head, Generalized Focal Loss, Varifocal Loss, Rank & Sort Loss) have shown that forcing these two loss terms to interact with each other in non-conventional ways creates a useful inductive bias and improves performance. Inspired by these works, we focus on the correlation between classification and localization and make two main contributions in this thesis: (i) We provide an analysis about the effects of correlation between classification and localization tasks in object detectors. We identify why correlation affects the performance of various NMS-based and NMS-free detectors, and we devise performance measures to evaluate the effect of correlation and use them to analyze common detectors. (ii) Motivated by our observations, e.g., that NMS-free detectors can also benefit from correlation, we propose Correlation Loss, a novel plug-in loss function that improves the performance of various object detectors by directly optimizing correlation coefficients: E.g., Correlation Loss on Sparse R-CNN, an NMS-free method, yields up to 1.6 AP points improvement, and our best model on Sparse R-CNN with ResNeXt-101 reaches 51.0 AP without test-time augmentation, outperforming all NMS-based and NMS-free detectors.

Supervisor(s)

Supervisor(s)

FEHMI KAHRAMAN

Date and Location

Date and Location

2022-08-18 10:00:00

Category

Category

MSc_Thesis