Graph-Based Point Tracker for 3D Object Tracking in Point Clouds
Authors: Minseong Park, Hongje Seong, Wonje Jang, Euntai Kim2053-2061
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on the KITTI tracking dataset show that GPT achieves state-of-the-art performance and can run in real-time. |
| Researcher Affiliation | Academia | Minseong Park, Hongje Seong, Wonje Jang, Euntai Kim* School of Electrical and Electronic Engineering, Yonsei University, Seoul, Korea {msp922, hjseong, jangwj1256, etkim}@yonsei.ac.kr |
| Pseudocode | No | The paper describes its method using text and diagrams, but it does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any statement about releasing source code or provide a link to a code repository. |
| Open Datasets | Yes | To validate our GPT, we use the KITTI tracking dataset as a benchmark (Geiger, Lenz, and Urtasun 2012). |
| Dataset Splits | Yes | As the ground truth (GT) of its test set is not available, we divide its training set into three sets, and use them for training, validation, and testing. The KITTI training set has 21 scenes. Following the settings in previous works (Giancola, Zarzar, and Ghanem 2019; Qi et al. 2020), we use scenes 0-16 for training, scenes 17-18 for validation and scenes 19-20 for testing. |
| Hardware Specification | No | The paper mentions running in real-time at '38 FPS on a single GPU', but it does not specify the model or type of GPU, or any other hardware components. |
| Software Dependencies | No | The paper states 'Point Net++ (Qi et al. 2017b) with three set abstraction (SA) layers is used as a backbone in GPT' but does not specify software versions for PointNet++ or any other dependencies like Python, PyTorch, TensorFlow, or CUDA. |
| Experiment Setup | Yes | In the three SA layers, the radii of their receptive fields are set to 0.3, 0.5, and 0.7 m, respectively. The size of the range query is set to 32, and the number of points is cut in half by random down-sampling in each SA layer. ... In GFA, template and search area seeds are transformed to nodes with Dnode = Dseed/2 = 128 feature dimension by φ and ϕ, respectively, and the bipartite graph G is constructed with k = 16. ... L = 64 clusters with the radius R = 0.3 m and 16 query points are generated. ... When we train our GPT, the relative weights of the loss are set to λsea-reg = 0.9, λtmp-reg = 0.1, λcls = 0.2, λbox = 0.2, and λprop = 1.5. An Adam optimizer is used, the batch size is 48, and the learning rate is initially set to 0.001, which is reduced by a rate of 0.2 every 12 epochs. |