DDFlow: Learning Optical Flow with Unlabeled Data Distillation
Authors: Pengpeng Liu, Irwin King, Michael R. Lyu, Jia Xu8770-8777
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a rigorous evaluation on the challenging Flying Chairs, MPI Sintel, KITTI 2012 and 2015 benchmarks, and show that our approach significantly outperforms all existing unsupervised learning methods, while running at real time. |
| Researcher Affiliation | Collaboration | Pengpeng Liu, Irwin King, Michael R. Lyu, Jia Xu The Chinese University of Hong Kong, Shatin, N.T., Hong Kong Tencent AI Lab, Shenzhen, China {ppliu, king, lyu}@cse.cuhk.edu.hk, jiajxu@tencent.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | To ensure reproducibility and advance further innovations, we make our code and models publicly available our our project website. |
| Open Datasets | Yes | We evaluate DDFlow on standard optical flow benchmarks including Flying Chairs (Dosovitskiy et al. 2015), MPI Sintel (Butler et al. 2012), KITTI 2012(Geiger, Lenz, and Urtasun 2012), and KITTI 2015 (Menze and Geiger 2015). |
| Dataset Splits | No | The paper mentions using 'training set' and 'test set' for standard benchmarks (Flying Chairs, MPI Sintel, KITTI), but does not explicitly detail the exact splits (percentages or counts) for training, validation, and testing or how the validation set was used/derived. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions using PWC-Net as a backbone and Adam optimizer but does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For all our experiments, we use the same network architecture and train our model using Adam optimizer (Kingma and Ba 2014) with β1 =0.9 and β2=0.999. For all datasets, we set batch size as 4. For all individual experiments, we use a initial learning rate of 1e-4, and it decays half every 50k iterations. For data augmentation, we only use random cropping, random flipping, and random channel swapping. |