Correlation Loss: Enforcing Correlation Between Classification and Localization in Object Detection

Kahraman, Fehmi
Object detectors are conventionally trained by a weighted sum of classification and localization losses. Recent studies (e.g., predicting IoU with an auxiliary head, Gen eralized Focal 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 be tween 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 perfor mance 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 detec tors. (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 im proves the performance of various object detectors by directly optimizing correlation coefficients: E.g., Correlation Loss on Sparse R-CNN, an NMS-free method, yields 1.6 AP gain on COCO dataset. Our best model on Sparse R-CNN reaches 51.0 AP without test-time augmentation on COCO test-dev, reaching state-of-the-art.


Scale invariant representation of 2 5D data
AKAGUNDUZ, Erdem; ULUSOY PARNAS, İLKAY; BOZKURT, Nesli; Halıcı, Uğur (2007-06-13)
In this paper, a scale and orientation invariant feature representation for 2.5D objects is introduced, which may be used to classify, detect and recognize objects even under the cases of cluttering and/or occlusion. With this representation a 2.5D object is defined by an attributed graph structure, in which the nodes are the pit and peak regions on the surface. The attributes of the graph are the scales, positions and the normals of these pits and peaks. In order to detect these regions a "peakness" (or pi...
Posterior Cram'er-Rao Lower Bounds for Extended Target Tracking with Random Matrices
Sarıtaş, Elif; Orguner, Umut (2016-07-08)
This paper presents posterior Cram'er-Rao lower bounds (PCRLB) for extended target tracking (ETT) when the extent states of the targets are represented with random matrices. PCRLB recursions are derived for kinematic and extent states taking complicated expectations involving Wishart and inverse Wishart distributions. For some analytically intractable expectations, Monte Carlo integration is used. The bounds for the semi-major and minor axes of the extent ellipsoid are obtained as well as those for the exte...
Correlation distribution of a sequence family generalizing some sequences of trachtenberg
Özbudak, Ferruh (2021-08-01)
In this paper, we give a classification of a sequence family, over arbitrary characteristic, adding linear trace terms to the function g(x) = Tr(x(d)), where d = p(2k) - p(k) + 1, first introduced by Trachtenberg. The family has p(n) + 1 cyclically distinct sequences with period p(n) - 1. We compute the exact correlation distribution of the function g(x) with linear m-sequences and amongst themselves. The cross-correlation values are obtained as C-i,C-j(tau) is an element of {-1, -1 +/- p(n+e/2), -1 + p(n)}.
Multisource region attention network for fine-grained object recognition in remote sensing imagery
Sümbül, Gencer; Cinbiş, Ramazan Gökberk; Aksoy, Selim (Institute of Electrical and Electronics Engineers (IEEE), 2019-07)
Fine-grained object recognition concerns the identification of the type of an object among a large number of closely related subcategories. Multisource data analysis that aims to leverage the complementary spectral, spatial, and structural information embedded in different sources is a promising direction toward solving the fine-grained recognition problem that involves low between-class variance, small training set sizes for rare classes, and class imbalance. However, the common assumption of coregistered ...
Covariance Matrix Estimation of Texture Correlated Compound-Gaussian Vectors for Adaptive Radar Detection
Candan, Çağatay; Pascal, Frederic (2022-01-01)
Covariance matrix estimation of compound-Gaussian vectors with texture-correlation (spatial correlation for the adaptive radar detectors) is examined. The texture parameters are treated as hidden random parameters whose statistical description is given by a Markov chain. States of the chain represent the value of texture coefficient and the transition probabilities establish the correlation in the texture sequence. An Expectation-Maximization (EM) method based covariance matrix estimation solution is given ...
Citation Formats
F. Kahraman, “Correlation Loss: Enforcing Correlation Between Classification and Localization in Object Detection,” M.S. - Master of Science, Middle East Technical University, 2022.