Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation
Authors: Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, Tiejun Huang525-533
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on the Multi Vehicle Stereo Event Camera dataset (MVSEC) (Zhu et al. 2018a) and demonstrate its superior generalization ability in different scenarios. Results show that STE-Flow Net outperforms all the existing state-of-the-art methods (Zhu et al. 2019; Lee et al. 2020). |
| Researcher Affiliation | Academia | Peking University {ziluo, gtx, rqxiong, yuzf12, tjhuang}@pku.edu.cn, {ruizhao, jyzhang}@stu.pku.edu.cn |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing the source code or a direct link to a code repository for the methodology described. |
| Open Datasets | Yes | The MVSEC dataset (Zhu et al. 2018a) is used for training and evaluating our model since it is designed for the development of visual perception algorithms for event-based cameras. |
| Dataset Splits | Yes | To provide fair comparisons with prior works (Zhu et al. 2019; Lee et al. 2020; Zhu et al. 2018b), we only use the outdoor day2 sequence to train the models. Indoor flying1, indoor flying2, indoor flying3, and outdoor day1 sequences are for evaluation only. |
| Hardware Specification | No | No specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running the experiments were found in the paper. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer (Kingma and Ba 2014)' but does not provide specific version numbers for software dependencies like programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | When it comes to dt = 1 case, the model is trained for 40 epochs. The number of event images Nframe that summarizes input event sequence is set to 5, and weight factor λ for the smoothness loss is set to 10. The initial learning rate is 4e 4 and scaled by 0.7 after 5, 10, and 20 epochs. As for dt = 4 case, the model is trained for 15 epochs. Nframe is set to 20 and λ is set to 10. In addition, the initial learning rate is 4e 4 with the same scaled strategy. Note that we use Adam optimizer (Kingma and Ba 2014) with minibatch size of 16 and the number of iteration for IRR Nirr is set to 3 in both cases. |