From ViT Features to Training-free Video Object Segmentation via Streaming-data Mixture Models

Authors: Roy Uziel, Or Dinari, Oren Freifeld

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the efficacy of the method on key benchmarks: the DAVIS-2017 and You Tube-VOS 2018 validation datasets. We report the results on two widely-used SVOS benchmarks: You Tube-VOS [46] and DAVIS 2017 [34]. We performed an ablation study (Fig. 6) to analyze the influence of different parts of the method on the performance.
Researcher Affiliation Academia Roy Uziel Ben-Gurion University of the Negev, Israel uzielr@post.bgu.ac.il Or Dinari Ben-Gurion University of the Negev, Israel dinari@post.bgu.ac.il Oren Freifeld Ben-Gurion University of the Negev, Israel orenfr@cs.bgu.ac.il
Pseudocode No The paper describes algorithmic steps but does not include structured pseudocode or clearly labeled algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/BGUCS-VIL/Training-Free-VOS.
Open Datasets Yes We report the results on two widely-used SVOS benchmarks: You Tube-VOS [46] and DAVIS 2017 [34].
Dataset Splits Yes We demonstrate the efficacy of the method on key benchmarks: the DAVIS-2017 and You Tube-VOS 2018 validation datasets. On the DAVIS-2017 validation set (Table 1)... We also evaluated our method on the DAVIS-2017 training and test-dev sets, collectively constituting 90 additional sequences.
Hardware Specification Yes Table 4: FPS across resolutions. Comparison on Tesla V100-32GB, excluding feature extraction.
Software Dependencies No Appendix B states 'We implemented our solution in PyTorch.' but does not specify a version number or other software dependencies with their versions.
Experiment Setup Yes Our baseline configuration (the centered one) is: S = 10, λ = 0.33, wρ = 15.