Learning Optical Flow with Adaptive Graph Reasoning
Authors: Ao Luo, Fan Yang, Kunming Luo, Xin Li, Haoqiang Fan, Shuaicheng Liu1890-1898
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On both Sintel clean and final passes, our AGFlow achieves the best accuracy with EPE of 1.43 and 2.47 pixels, outperforming state-of-the-art approaches by 11.2% and 13.6%, respectively. Code is publicly available at https://github.com/ megvii-research/AGFlow. We conduct extensive experiments on two standard datasets, i.e., MPI-Sintel (Butler et al. 2012) and KITTI 2015 (Menze and Geiger 2015). |
| Researcher Affiliation | Collaboration | Ao Luo1, Fan Yang2, Kunming Luo1, Xin Li2, Haoqiang Fan1, Shuaicheng Liu3,1* 1Megvii Technology 2Group 42 3University of Electronic Science and Technology of China |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. It provides architectural diagrams and mathematical formulations instead. |
| Open Source Code | Yes | Code is publicly available at https://github.com/ megvii-research/AGFlow. |
| Open Datasets | Yes | The model is pretrained on Flying Chairs (Dosovitskiy et al. 2015) for 180k iterations and then on Flying Things (Mayer et al. 2016) for 180k iterations. After that, we fine-tune the model on combined data from Sintel (Butler et al. 2012), KITTI-2015 (Menze and Geiger 2015), and HD1K (Kondermann et al. 2016) for 180k iterations |
| Dataset Splits | Yes | We conduct extensive experiments on two standard datasets, i.e., MPI-Sintel (Butler et al. 2012) and KITTI 2015 (Menze and Geiger 2015). We follow prior works (Teed and Deng 2020; Jiang et al. 2021b) to utilize two standard evaluation metrics, i.e., average end-point error (EPE) and the percentage of erroneous pixels > 3 pixels (F1-all), to evaluate the performance of predicted optical flow. On Sintel (val) and KITTI-15 (val) in Table 1. |
| Hardware Specification | Yes | Our model is trained on 2 NVIDIA Ge Force GTX 2080Ti GPUs, and the batch size is set to 8 for better leveraging the GPU memory. |
| Software Dependencies | No | The implementation of our approach is based on Py Torch toolbox. This statement mentions PyTorch but does not provide a specific version number, nor does it list other software dependencies with version numbers. |
| Experiment Setup | Yes | In our model, we set the number of context and motion nodes K to 128. The state updating iterations t are set to 2 and 1 for context and motion graph, respectively. During training, we follow prior works (Teed and Deng 2020; Jiang et al. 2021a) to adopt Adam W optimizer with one-cycle learning rate policy... The model is pretrained on Flying Chairs (Dosovitskiy et al. 2015) for 180k iterations and then on Flying Things (Mayer et al. 2016) for 180k iterations. After that, we fine-tune the model on combined data from Sintel (Butler et al. 2012), KITTI-2015 (Menze and Geiger 2015), and HD1K (Kondermann et al. 2016) for 180k iterations... Finally, additional 50k iterations of finetuning are performed on KITTI-2015... the batch size is set to 8 |