Access to Image metadata to create realistic train/test splits

[seyuboglu]

Cool task – thanks for putting together this resource.

Many of the images in the dataset are near duplicates of one another – presumably because the KITTI images were taken from videos:

Splitting randomly, as done in https://github.com/harshilpatel312/KITTI-distance-estimation/blob/3a79107d9a07f8a566ca36eb6318a22ae78116f1/generate-depth-annotations.py, would make for an unrealistic task. Images taken seconds apart could straddle the train/test split.

Ideally, we would split the dataset such that photos from the same drive or same geographic region do not straddle the splits. Does anyone know if this metadata available in the Kitti benchmark?