A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object Detection

Download

index.pdf

Date

2020-12-06

Author

Öksüz, Kemal
Çam, Barış Can
Akbaş, Emre
Kalkan, Sinan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

140
views

37
downloads

We propose average Localisation-Recall-Precision (aLRP), a unified, bounded, balanced and ranking-based loss function for both classification and localisation tasks in object detection. aLRP extends the Localisation-Recall-Precision (LRP) performance metric (Oksuz et al., 2018) inspired from how Average Precision (AP) Loss extends precision to a ranking-based loss function for classification (Chen et al., 2020). aLRP has the following distinct advantages: (i) aLRP is the first ranking-based loss function for both classification and localisation tasks. (ii) Thanks to using ranking for both tasks, aLRP naturally enforces high-quality localisation for high-precision classification. (iii) aLRP provides provable balance between positives and negatives. (iv) Compared to on average ~6 hyperparameters in the loss functions of state-of-the-art detectors, aLRP Loss has only one hyperparameter, which we did not tune in practice. On the COCO dataset, aLRP Loss improves its ranking-based predecessor, AP Loss, up to around 5 AP points, achieves 48.9 AP without test time augmentation and outperforms all one-stage detectors. Code available at: https://github.com/kemaloksuz/aLRPLoss .

URI

https://hdl.handle.net/11511/92442

Conference Name

2020 Conference on Neural Information Processing Systems

Collections

Department of Computer Engineering, Conference / Seminar

Suggestions

OpenMETU
Core

A probabilistic sparse skeleton based object detection Altinoklu, Burak; Ulusoy, İlkay; Tarı, Zehra Sibel (Elsevier BV, 2016-11) We present a Markov Random Field (MRF) based skeleton model for object shape and employ it in a probabilistic chamfer-matching framework for shape based object detection. Given an object category, shape hypotheses are generated from a set of sparse (coarse) skeletons guided by suitably defined unary and binary potentials at and between shape parts. The Markov framework assures that the generated samples properly reflect the observed or desired shape variability. As the model employs a sparsely sampled skele...
An mpeg-7 video database system for content-based management and retrieval Çelik, Çiğdem; Çiçekli, Fehime Nihan; Department of Computer Engineering (2005) A video data model that allows efficient and effective representation and querying of spatio-temporal properties of objects has been previously developed. The data model is focused on the semantic content of video streams. Objects, events, activities performed by objects are the main interests of the model. The model supports fuzzy spatial queries including querying spatial relationships between objects and querying the trajectories of objects. In this thesis, this work is used as a basis for the developmen...
A probabilistic multiple criteria sorting approach based on distance functions ÇELİK, BİLGE; Karasakal, Esra; İyigün, Cem (2015-05-01) In this paper, a new probabilistic distance based sorting (PDIS) method is developed for multiple criteria sorting problems. The distance to the ideal point is used as a criteria disaggregation function to determine the values of alternatives. These values are used to sort alternatives into the predefined classes. The method also calculates probabilities that each alternative belong to the predefined classes in order to handle alternative optimal solutions. It is applied to five data sets and its performanc...
A Fast and Automatically Paired 2-Dimensional Direction-of-Arrival Estimation Using Arbitrary Array Geometry Filik, T.; Tuncer, Temel Engin (2009-04-11) A new approach is proposed for two-dimensional (2-D) direction-of-arrival (DOA) estimation with arbitrary array geometries, which is based on array interpolation. The method provides automatically paired source azimuth and elevation angle estimates. Furthermore it is possible to estimate D sources with D + 1 sensor 2-D array interpolation errors are minimized by using Wiener formulation. Proposed method is applied to the two planar arrays; uniform circular array (UCA) and uniform isotropic (IU) V-shaped arr...
A method in model updating using Miscorrelation Index sensitivity Kozak, Mustafa Tuğrul; Öztürk, Murat; Özgüven, Hasan Nevzat (Elsevier BV, 2009-08-01) This paper presents a new model updating method based on minimization of an index called Miscorrelation Index (MCI), which is introduced to localize the coordinates carrying error in a finite element (FE) model. MCI can be calculated from measured frequency response functions (FRFs) and dynamic stiffness matrix of the FE model for each coordinate as a function of frequency. Nonzero numerical values for MCI of a coordinate indicate errors in one or more elements of the system matrices corresponding to this c...

Citation Formats

K. Öksüz, B. C. Çam, E. Akbaş, and S. Kalkan, “A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object Detection,” Montreal, Kanada, 2020, vol. 33, Accessed: 00, 2021. [Online]. Available: https://hdl.handle.net/11511/92442.