nordland subset

[xjh19971]

Following [1-2], this script provides a subset of nordland dataset, which contains 2760 query images instead of 27k query images. Meanwhile, the subset requires THRESHOLD_FRAMES=1 instead of 10.

[1] Zaffar, M., Garg, S., Milford, M. et al. VPR-Bench: An Open-Source Visual Place Recognition Evaluation Framework with Quantifiable Viewpoint and Appearance Change. Int J Comput Vis 129, 2136–2174 (2021). https://doi.org/10.1007/s11263-021-01469-5
[2] Ali-bey, A., Chaib-draa, B., & Giguère, P. (2024). BoQ: A Place is Worth a Bag of Learnable Queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 17794-17803).

[gmberton]

Hi, sorry for the late response. Did you try to test some VPR model on the dataset created by this code? Shouldn't the DISTANCE_BETWEEN_FRAMES be different between database and queries?

[xjh19971]

No, I don't modify the distance_between_frames. Instead, I adjust the sampling frequency between the database and query sets: database images are sampled for every frame, while query images are sampled every 10 frames. Feel free to change this if you find a better way.

To ensure correctness, I compare the generated dataset with the one provided in the SALAD repository (https://github.com/serizba/salad?tab=readme-ov-file). Specifically, I cross-check the first 10 frames with the reference dataset available here: https://surfdrive.surf.nl/files/index.php/s/sbZRXzYe3l0v67W. My comparison shows that the first 10 frames in my dataset match those from this repo.

One additional observation I made is that the two datasets appear to use different image upsampling/downsampling methods. Initially, I wrote a script to compare the datasets pixel-by-pixel, but I noticed slight differences in pixel values. As a result, I had to verify the content manually.