DistilVPR: Cross-Modal Knowledge Distillation for Visual Place Recognition
Authors: Sijie Wang, Rui She, Qiyu Kang, Xingchao Jian, Kai Zhao, Yang Song, Wee Peng Tay
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments demonstrate that our proposed pipeline achieves state-of-the-art performance compared to other distillation baselines. We also conduct necessary ablation studies to show design effectiveness. Through extensive experiments, we showcase the remarkable performance of Distil VPR when compared to previous KD baselines. Our approach achieves state-of-the-art (SOTA) performance in the task of cross-modal distillation for VPR. Furthermore, we rigorously investigate our design through vital ablation studies, providing empirical evidence of the efficacy of our proposed methodology. |
| Researcher Affiliation | Academia | Sijie Wang*, Rui She*, Qiyu Kang , Xingchao Jian, Kai Zhao, Yang Song, Wee Peng Tay Nanyang Technological University {wang1679, rui.she, qiyu.kang, xingchao001}@ntu.edu.sg, yang.song@connect.polyu.hk |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. It provides mathematical formulations and descriptions in text. |
| Open Source Code | Yes | The code is released at: https://github.com/sijieaaa/Distil VPR |
| Open Datasets | Yes | Oxford Robot Car. The Oxford Robot Car dataset (Maddern et al. 2017) is a large-scale autonomous driving dataset... Boreas. The Boreas dataset (Burnett et al. 2022) is gathered by conducting multiple drives along a consistent route over one year... |
| Dataset Splits | No | The paper mentions using datasets for training, but does not explicitly provide details about specific training, validation, and test splits (e.g., percentages or sample counts for each split). |
| Hardware Specification | Yes | All experiments are conducted on an A100 GPU. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer' but does not specify version numbers for any software dependencies (e.g., Python, PyTorch, TensorFlow, CUDA). |
| Experiment Setup | Yes | We use the Adam optimizer to train both teachers and students. The learning rate is set as 1e 4 and 1e 3 for the image branch and the point cloud branch respectively. Both teacher models and student models are trained for 60 epochs with 128 batch size. |