BAR — A Reinforcement Learning Agent for Bounding-Box Automated Refinement

Authors: Morgane Ayle, Jimmy Tekli, Julia El-Zini, Boulos El-Asmar, Mariette Awad2561-2568

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results on a car industry-related dataset and on the PASCAL VOC dataset show a consistent increase of up to 0.28 in the Intersection-over-Union of bounding-boxes with their desired ground-truths, while saving 30%-82% of human intervention time in either correcting or re-drawing inaccurate proposals.
Researcher Affiliation Collaboration 1Maroun Semaan Faculty of Engineering and Architecture, American University of Beirut, Lebanon 2Universite de Franche Comtee, Belfort, France 3BMW Group, Munich, Germany
Pseudocode No The paper describes algorithmic steps in prose but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes 2Code available at https://gitlab.com/Morgane Ayle/bar
Open Datasets Yes For a fair comparison with the literature, BAR-DRL is additionally evaluated on the aeroplane class of the publicly available PASCAL VOC dataset (Everingham et al. 2010).
Dataset Splits Yes There are approximately 800 available images for each class, 200 of which are used for testing. and BAR-DRL is trained on the latter [PASCAL 2007 trainval], with 309 training examples, β = 0.6 and c1 = 0.2. The pipeline is tested on PASCAL 2007 test data
Hardware Specification Yes The experiments were run on a Ge Force GTX 1080 with 11,000 Mi B memory and Cuda version 10.0.
Software Dependencies Yes The experiments were run on a Ge Force GTX 1080 with 11,000 Mi B memory and Cuda version 10.0.
Experiment Setup Yes The neural network consists of two fully-connected layers of 500 neurons each with Relu activation and random normal initialization, and an output layer of 9 neurons with linear activation. Mean square error loss with Adam optimizer and learning rate of 0.001 are used, and the discount factor γ for the Q-function is set to 0.90. BAR-DRL is trained for 80 epochs using an experience replay with maximum memory length of 2000, and epsilon-greedy exploration with ϵ = 1.0 initially and linearly annealing to 0.1 for 15 epochs.