EGSST: Event-based Graph Spatiotemporal Sensitive Transformer for Object Detection

Authors: Sheng Wu, Hang Sheng, Hui Feng, Bo Hu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we introduce the two datasets utilized, the evaluation metrics, and the implementation details of our models. We train the baseline model, EGSST-B, and the extended model, EGSST-E, and compare their performance with other state-of-the-art models applied to both datasets. Detailed ablation studies are then performed to assess the impact of various components of our models.
Researcher Affiliation Academia Sheng Wu1 Hang Sheng1 Hui Feng1,2 Bo Hu1,2 1 School of Information Science and Technology, Fudan University 2 State Key Laboratory of Integrated Chips and Systems, Fudan University
Pseudocode No The paper describes its method in detail but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The source code can be found at: EGSST.
Open Datasets Yes Two complex event camera datasets from traffic scenarios are employed in the experiments: the Gen1 Automotive Detection Dataset [46] and the 1 Megapixel Automotive Detection Dataset [9].
Dataset Splits No The paper mentions using datasets for training and testing, and sets a "training batch size", but it does not explicitly provide details on the training/validation/test splits, such as percentages, sample counts, or references to predefined splits.
Hardware Specification Yes The models are trained on RTX3090 GPUs using the Lightning framework...we conducte additional tests on the T4 GPU, which has performance comparable to the Titan Xp and RTX 1080Ti.
Software Dependencies Yes The framework proposed in this study is developed using Python 3.9 and Py Torch 2.0, with graph processing powered by the advanced Py Torch Geometric library [48].
Experiment Setup Yes We employ the Adam optimizer [49] coupled with the One Cycle learning rate schedule [50], which includes 100 warm-up iterations followed by cosine decay starting from the maximum learning rate. The training batch size is set at 8, with an initial learning rate of 1e-6.