An Analytical Solution to Gauss-Newton Loss for Direct Image Alignment

Authors: Sergei Solonets, Daniil Sinitsyn, Lukas Von Stumberg, Nikita Araslanov, Daniel Cremers

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on two most popular datasets for large-scale image localization, namely the Aachen Day-Night dataset (Sattler et al., 2018), extended CMU seasons (Toft et al., 2022) and 7Scenes dataset (Shotton et al., 2013). Table 1 presents the results for camera localization on Aachen Day-Night (Sattler et al., 2018) and the extended CMU Seasons (Toft et al., 2022).
Researcher Affiliation Collaboration Sergei Solonets1,2, Daniil Sinitsyn1,2, Lukas von Stumberg3, Nikita Araslanov1,2 Daniel Cremers1,2 1 Technical University of Munich 2 Munich Center for Machine Learning 3 Valve Software
Pseudocode Yes Algorithm 1: Image alignment.
Open Source Code Yes Project code: https://github.com/tum-vision/gn_loss_analytical. To facilitate reproducibility in future research, we also publicly release our code.
Open Datasets Yes We evaluate our approach on two most popular datasets for large-scale image localization, namely the Aachen Day-Night dataset (Sattler et al., 2018), extended CMU seasons (Toft et al., 2022) and 7Scenes dataset (Shotton et al., 2013).
Dataset Splits No The paper evaluates its approach on established datasets and uses self-supervised feature descriptors like Super Point. However, it does not explicitly provide training/validation/test dataset splits for its own method or the pre-trained feature extractors, which are used as components.
Hardware Specification Yes On average, the code takes approximately 6 seconds and 10 seconds per alignment on CMU and Aachen, respectively, on a single NVIDIA A4000.
Software Dependencies No The paper states 'We implement our approach in Py Torch (Paszke et al., 2019)' but does not specify the version number of PyTorch or any other software dependencies.
Experiment Setup Yes We initialize p( ) with a truncated uniform distribution of a fixed radius around all interest points. The radius decreases from 50% of the image diagonal to 5% in the first 30 iterations. Afterward, the scheduler switches to the normal distribution around each interest point with standard deviation σ. Initially, we define σ such that 99% of the distribution covers 10% of the image around the point, and we decrease the coverage ratio to 1%. In our experiments with Super Point, we set the threshold value to 0.4.