Merge or Not? Learning to Group Faces via Imitation Learning

Authors: Yue He, Kaidi Cao, Cheng Li, Chen Loy

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on three benchmark datasets show that our framework outperforms unsupervised and supervised baselines.
Researcher Affiliation Collaboration 1Sense Time Group Limited, {heyue,caokaidi,chengli}@sensetime.com 2The Chinese University of Hong Kong, ccloy@ie.cuhk.edu.hk
Pseudocode Yes Algorithm 1: Reward function learning via IRL.
Open Source Code Yes Our codes and data are released1 to fa- 1https://github.com/bj80heyue/Learning-to-Group cilitate future studies.
Open Datasets Yes We employ 2, 000 albums simulated from MS-Celeb1M (Guo et al. 2016) of 80k identities as our training source and generalize it to various test data below. ... LFW (Huang et al. 2007), MS-Celeb-1M (Guo et al. 2016), and PFW (Sengupta et al. 2016) ... ACCIO dataset (Ghaleb et al. 2015) ... Our codes and data are released1 to fa- 1https://github.com/bj80heyue/Learning-to-Group cilitate future studies.
Dataset Splits No The paper states using "2,000 albums simulated from MS-Celeb1M" as a training source and evaluating on "LFW-Album", "ACCIO Dataset", and "Grouping Face in the Wild (GFW)". While it specifies quantities for some datasets (e.g., "20 albums" for LFW-Album, "3243 tracklets" for ACCIO), it does not provide explicit numerical train/validation/test splits (e.g., percentages or sample counts for each split) within these datasets needed for reproducibility of data partitioning.
Hardware Specification No No specific hardware details (such as GPU/CPU models, memory, or processor types) used for running the experiments were mentioned in the paper.
Software Dependencies No The paper mentions models and algorithms such as Inception-v3, SVM, random forest regressor, and Faster-RCNN, but it does not provide specific version numbers for these or any other software dependencies (e.g., Python version, library versions) that would be needed for replication.
Experiment Setup Yes We set β = 0.8 in Eqn. (3) to balance the scales of shortand long-term rewards. We fixed the number of faces η = 5 to form the similarity and quality features. ... Specifically, we set γ = 0 in Eqn. (2) and β = 0 in Eqn. (3). ... Specifically, we set γ = 0.9 in Eqn. (2) and β = 0.8 in Eqn. (3). ... The three layers of the Siamese network have 256, 64, 64 hidden neurons, respectively. A contrastive loss is used for training.