Scale-Free Image Keypoints Using Differentiable Persistent Homology
Authors: Giovanni Barbarani, Francesco Vaccarino, Gabriele Trivigno, Marco Guerra, Gabriele Berton, Carlo Masone
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our detector, Morse Det, is the first topology-based learning model for feature detection, which achieves competitive performance in keypoint repeatability... We assessed the capability of our method to predict repeatable keypoints using the well-established HPatches benchmark (Balntas et al., 2017). The results are shown in the tab. 1... We conducted an ablation study on the effect of training with different values of α. The results regarding the repeatability of the HPatches viewpoint data split are reported in the tab. 3. |
| Researcher Affiliation | Academia | 1Department of Mathematical Sciences Giuseppe Luigi Lagrange , Politecnico di Torino, Italy 2Department of Control and Computer Engineering, Politecnico di Torino, Italy 3Institut Fourier, Universit e Grenoble Alpes, France. |
| Pseudocode | No | The paper describes the methods in prose and with figures, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our implementation and trained models have been publicly released1. 1https://github.com/gbarbarani/MorseDet |
| Open Datasets | Yes | For training our detector, we adopt WASF, the dataset released in (Revaud et al., 2019) to train R2D2, which provides homographic correspondences between pairs of images. Hyperparameter search and early stopping are performed on the validation split of Mega Depth (Li & Snavely, 2018) used in (De Tone et al., 2018). |
| Dataset Splits | Yes | Hyperparameter search and early stopping are performed on the validation split of Mega Depth (Li & Snavely, 2018) used in (De Tone et al., 2018). |
| Hardware Specification | Yes | The training process on a single TITAN X GPU with 12GB of VRAM concluded in approximately 10 hours until convergence. |
| Software Dependencies | No | The paper mentions "Adam W (Loshchilov & Hutter, 2019) as an optimizer" but does not provide specific version numbers for any software, libraries, or frameworks used (e.g., Python, PyTorch, TensorFlow, etc.). |
| Experiment Setup | Yes | The final hyperparameters configuration included α = 10, weight decay equal to 0.005, and repeatability threshold γ = 0.7 for inference. For training, we employed Adam W... with batches of 8 pairs of images with resolution 208 208. |