Space-Time Correspondence as a Contrastive Random Walk

Authors: Allan Jabri, Andrew Owens, Alexei Efros

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the learned representation on video label propagation tasks involving objects, keypoints, and semantic parts, by using it as a similarity metric. We also study the effects of edge dropout, training sequence length, and self-supervised adaptation at test-time. In addition to comparison with the state-of-the-art, we consider a baseline of label propagation with strong pre-trained features.
Researcher Affiliation Academia Allan A. Jabri UC Berkeley Andrew Owens University of Michigan Alexei A. Efros UC Berkeley
Pseudocode Yes Algorithm 1 provides complete pseudocode for the method.
Open Source Code No The paper does not contain an explicit statement about the release of source code for the described methodology or a direct link to a code repository.
Open Datasets Yes We train φ using the (unlabeled) videos from Kinetics400 [14], with Algorithm 1. We evaluate our model on DAVIS 2017 [83], a popular benchmark for video object segmentation... We consider pose tracking on the JHMDB benchmark... We consider the semantic part segmentation task of the Video Instance Parsing (VIP) benchmark [120]
Dataset Splits No The paper mentions using specific datasets (Kinetics400, DAVIS 2017, JHMDB, VIP) for training and evaluation, but it does not explicitly provide the train/test/validation split percentages or sample counts used for these datasets within the main text.
Hardware Specification No The paper mentions 'compute resources donated by NVIDIA' in the acknowledgments but does not provide specific details on the GPU models, CPU models, or other hardware specifications used for running the experiments.
Software Dependencies No The paper mentions a 'PyTorch-like style' for its pseudocode and references libraries and optimizers like ResNet-18 [42] and Adam [49], but it does not provide specific version numbers for PyTorch or any other software dependencies required to replicate the experiments.
Experiment Setup Yes We train φ using the (unlabeled) videos from Kinetics400 [14], with Algorithm 1. We used the Adam optimizer [49] for two million updates with a learning rate of 1 10 4. We use a temperature of = 0.07 in Equation 1, following [113] and resize frames to 256 256 (before extracting nodes, as above). Except when indicated otherwise, we report results with edge dropout rate 0.1 and a videos of length 10.