SMILEtrack: SiMIlarity LEarning for Occlusion-Aware Multiple Object Tracking

Authors: Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung-Hin So, Xin Li

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments were conducted on MOT17 and MOT20 benchmarks (Milan et al. 2016)
Researcher Affiliation Academia 1College of Artificial Intelligence and Green Energy, National Yang Ming Chiao Tung University, Taiwan 2Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan 3Department of Computer Science, University at Albany SUNY, USA 4The Chinese University of Hong Kong, China
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. It describes the method in text and uses figures/equations, but no structured code-like representation.
Open Source Code Yes Code is available at https://github.com/pingyang1117/SMILEtrack Official.
Open Datasets Yes Our experiments were conducted on MOT17 and MOT20 benchmarks (Milan et al. 2016), with additional training on datasets (Sch ops et al. 2017; Doll ar et al. 2009; Milan et al. 2016; Zhang, Benenson, and Schiele 2017; Shao et al. 2018; Ess et al. 2008; Xiao et al. 2017; Zheng et al. 2017). For re-ID models, datasets providing both bounding box location and identity information, such as Cal Tech (Doll ar et al. 2009), PRW (Zheng et al. 2017), and CUHK-SYSU (Xiao et al. 2017), were used. Our detector was initialized on the COCO dataset (Lin et al. 2014) and finetuned on MOT datasets.
Dataset Splits Yes Table 4 shows the effects of different patch layouts on performance improvements evaluated on the MOT17 val set (Milan et al. 2016).
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions employing an "SGD optimizer with cosine annealing" and using a detector, but does not provide specific version numbers for any software components or libraries.
Experiment Setup No Our detector was initialized on the COCO dataset (Lin et al. 2014) and finetuned on MOT datasets, employing data augmentation and an SGD optimizer with cosine annealing. The SMC module introduced a GATE function to manage new tracklets, with key parameters assessed in an ablation study. While some details are mentioned, specific hyperparameters like learning rate, batch size, or exact training schedules are not provided.