Soft Contrastive Learning for Visual Localization
Authors: Janine Thoma, Danda Pani Paudel, Luc V. Gool
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on four large-scale benchmark datasets demonstrate the superiority of our soft contrastive learning over the state-of-the-art method for retrieval-based visual localization. 5 Experimental Evaluation |
| Researcher Affiliation | Academia | Janine Thoma1 Danda Pani Paudel1 Luc Van Gool1,2 1 Computer Vision Lab ETH Zurich, Switzerland 2 VISICS, ESAT/PSI KU Leuven, Belgium |
| Pseudocode | No | The paper describes mathematical formulations and processes, but it does not contain a dedicated 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | Our dataset and source code is made publicly available at https://github.com/janinethoma/ soft_contrastive_learning. |
| Open Datasets | Yes | We conduct experiments with models trained on three different publicly available real world datasets Image Net [44], Pittsburgh (Pitts250k) [45], and Oxford Robot Car [46]. |
| Dataset Splits | No | The paper states, 'The training is stopped once the validation performance stops improving,' indicating validation was used. However, it does not provide specific details on the validation dataset split (e.g., percentages or counts). |
| Hardware Specification | No | The paper only states: 'All models are trained on a single Nvidia GPU.' This is not specific enough to identify the hardware model or other specifications. |
| Software Dependencies | No | The paper mentions using 'VGG-16' and 'Net VLAD [...] as implemented by [48],' but it does not specify software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, or specific library versions). |
| Experiment Setup | Yes | We use the training parameters of [16], reducing the learning rate to 0.000001. Each of our training tuples consists of 1 anchor, 12 close neighbors and 12 further away images. Images are down-scaled such that the largest side has a length of 240px. ... For our method and for log-ratio loss, we set r1 and r2 to 15m. For all other methods we set r1 to 10m and r2 to 25m in accordance with [16, 31, 30]. Similar to [16] we use hard negative mining with a mining cache size of 1000, which is updated every 250 steps. |