Unsupervised Hierarchical Domain Adaptation for Adverse Weather Optical Flow
Authors: Hanyu Zhou, Yi Chang, Gang Chen, Luxin Yan
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive quantitative and qualitative experiments have been performed to verify the superiority of the proposed method. In Table 1, we show the quantitative comparison on the synthetic heavy and light Rain-KITTI2015 dataset. In Fig. 3 and 4, we show the visual comparison results on real GOF and collected datasets, respectively. In Table 2, the quantitative results on GOF dataset further verify the superiority of the proposed method. To illustrate the effectiveness of the hierarchical adaptation architecture, in Fig. 5, we show the optical flow estimation of different architectures and visualize their low-dimensional distributions via t-SNE. In Table 3, we show how motion-boundary adaptation losses contribute to the final result. |
| Researcher Affiliation | Academia | 1 National Key Laboratory of Science and Technology on Multispectral Information Processing, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology 2 School of Computer Science and Engineering, Sun Yat-sen University |
| Pseudocode | No | The paper describes the architecture and method steps but does not include a formal pseudocode block or algorithm listing. |
| Open Source Code | No | The paper does not provide any statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | Datasets. We validate the performance on one synthetic and two real degraded datasets. We simulate the synthetic rain with different densities of rain (e.g., heavy rain and light rain) on KITTI2015 by the software Adobe After Effects. GOF (Li, Luo, and Liu 2021) is a real dataset with 1000 images for training and 120 images for testing. |
| Dataset Splits | No | The paper mentions 1000 images for training and 120 for testing for GOF, and 1200 for training and 200 for testing for the Youtube dataset. However, it does not specify explicit validation splits (e.g., percentages or counts) or reference predefined splits with citations for reproducibility. |
| Hardware Specification | Yes | Our method is implemented on the Tensorflow platform with two NVIDIA RTX 2080Ti GPUs within three days. |
| Software Dependencies | No | The paper mentions "Tensorflow platform" and uses "PWC-Net" as a backbone, but it does not specify version numbers for these or any other software dependencies, making it irreproducible based on the provided information. |
| Experiment Setup | Yes | We empirically set the parameters {α, β, γ1, γ2, δ, ρ1, ρ2} = {1, 1, 1, 1, 1, 0.1, 0.1}. Then we update the optical flow encoders and decoder via Lalign feat , Lconsis flow and Lprior flow with 200k iterators and 0.0002 learning rate, where we initialize the optical flow network of clean domain with clean KITTI2015 dataset (Menze and Geiger 2015) via Lprior flow only. |