Learning Graph-embedded Key-event Back-tracing for Object Tracking in Event Clouds
Authors: Zhiyu Zhu, Junhui Hou, Xianqiang Lyu
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on both synthetic and real event datasets demonstrate the superiority of the proposed framework over state-of-the-art methods in terms of both the tracking accuracy and speed. |
| Researcher Affiliation | Academia | Department of Computer Science, City University of Hong Kong |
| Pseudocode | No | The paper describes the proposed method in detail using text and diagrams, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | The code is publicly available at https://github.com/ZHU-Zhiyu/Event-tracking. |
| Open Datasets | Yes | We evaluated the proposed method with both real and synthetic event datasets. The real event dataset FE108 (MIT) [51]... Evt-La SOT (Apache-2.0) [12] via Vid2E[15]. |
| Dataset Splits | Yes | There are 59K event clouds for validation and 140K event clouds for training. |
| Hardware Specification | Yes | We implemented the framework with Pytorch on Ubuntu 18.04 and trained it with 4 NVIDIA-3090 and a batch size of 22 samples in each GPU , about 300K iterations. |
| Software Dependencies | No | The paper mentions 'Pytorch on Ubuntu 18.04' but does not specify version numbers for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | To train our framework in a batch manner, we sampled 10K event points for each event cloud. We set the value of t as the interval of annotations, i.e., 25 50ms for the FE108 dataset according to its 20/40 FPS annotations. We implemented the framework with Pytorch on Ubuntu 18.04 and trained it with 4 NVIDIA-3090 and a batch size of 22 samples in each GPU , about 300K iterations. We adopted the Adam optimizer [24] with the weight decay of 1e 5 and the learning rate of 1e 5 (resp. 1e 4) for the backbone (resp. the two target likelihood prediction and object proposal modules). We set the spatial search region of the current event cloud as 1.2 the size of the previous proposal. |