Tracklet Self-Supervised Learning for Unsupervised Person Re-Identification

Authors: Guile Wu, Xiatian Zhu, Shaogang Gong12362-12369

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superiority of TSSL over a wide variety of the stateof-the-art alternative methods on four large-scale person reid benchmarks, including Market-1501, Duke MTMC-Re ID, MARS and Duke MTMC-Video Re ID.
Researcher Affiliation Collaboration 1Queen Mary University of London, 2Vision Semantics Limited
Pseudocode Yes Algorithm 1 Tracklet Self-Supervised Learning.
Open Source Code No The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We aim at optimising a feature embedding space for both image and video unsupervised re-id, so we also evaluated both image (Market-1501 (Zheng et al. 2015) and Duke MTMCRe ID (Ristani et al. 2016; Zheng, Zheng, and Yang 2017)) and video (MARS (Zheng et al. 2016) and Duke MTMCVideo Re ID (Ristani et al. 2016; Wu et al. 2018a)) datasets.
Dataset Splits No The paper does not explicitly mention a validation split, nor does it detail the percentages or counts for training, validation, and test sets. It mentions a "maximal training epoch" but no specific validation process for hyperparameter tuning or early stopping.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions "We used Res Net-50 (He et al. 2016) (pre-trained on Image Net) as the feature embedding network." but does not provide version numbers for ResNet or any other software libraries or dependencies.
Experiment Setup Yes We empirically set α = 2 for Eq. (4), η = 0.5 for Eq. (6), λ = 0.1 and s = 10 for Eq. (5), τ = 0.1 for Eq. (7), δ = 0.05. We set Nk = 4 for cluster merging. The maximal training epoch was set to 20 for the first step and to 5 for the remaining steps. We used Stochastic Gradient Descent (SGD) as the optimiser with the initial learning rate at 0.01 for the backbone model and a decay of 0.1 after 15 training epochs.