Volumetric Correspondence Networks for Optical Flow

Authors: Gengshan Yang, Deva Ramanan

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our innovations dramatically improve accuracy over SOTA on standard benchmarks while being significantly easier to work with training converges in 7X fewer iterations
Researcher Affiliation Collaboration Gengshan Yang1 , Deva Ramanan1,2 1Carnegie Mellon University, 2Argo AI {gengshay, deva}@cs.cmu.edu
Pseudocode No Not found. The paper describes methods and components but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code will be available at github.com/gengshay-y/VCN.
Open Datasets Yes Number of training iterations is recorded for the pre-training stage on Flying Chairs and Flying Things, and (S) indicates sequential training on separate modules.
Dataset Splits No Number of training iterations is recorded for the pre-training stage on Flying Chairs and Flying Things, and (S) indicates sequential training on separate modules. As shown in Tab. 2, after the pretraining stage, ours-small achieves smaller end-point-error (EPE) than all methods on KITTI [9, 30]
Hardware Specification Yes The model is trained on a machine with 4 Titan X Pascal GPUs.
Software Dependencies No We build the model and re-implement the training pipeline of PWC-Net+ [39] using Pytorch.
Experiment Setup Yes We find correspondences with 9 9 search windows on a feature pyramid with stride {64, 32, 16, 8, 4}. We keep K = {16, 16, 16, 16, 12} hypotheses at each scale. To be noted, we are able to stably train the network with a larger learning rate (10 3 vs 10 4) and fewer iterations (140K vs 1200K on Flying Chairs and 80K vs 500K on Flying Things) compared to prior optical flow networks.