mapillary / mapillary_vistas

MVD Evaluation Scripts

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Inconsistency in config.json file and in annotated images

mitalbert opened this issue · comments

I'm sorry I'm writing here this issue related to dataset files as I couldn't find a feedback form on the website.

The download file (v1.1) includes a file named config.json which includes all categories and information about them. One category named "object--vehicle--on-rails" is supposedly an object and should have instances however the json file says "instances": false on that category. Is this is typo or am I missing something?

I tried to find instances of that object and found this annotation where rails are annotated as on-rails.
On the other hand, this annotation seems to be correct.

Hi @mitalbert!

Thanks for reporting the wrong annotation, that's definitely an error.

Regarding the instance flag for on-rails we decided to not make it instance-specific because we do not separate between different wagons and we probably (hopefully) never observed two trains on the same rail (or also in adjacent ones). Hence, there is not enough data to properly learn instance-separation and we turned it off all along.

So if on-rails category is not instance specific anymore, this means that the information concerning objects in blog post and in this paper is not correct for the database v1.1. Are there any changes other than this and if so, is there a post about these changes?

on-rails was never instance specific (not in v1.0 and not in v1.1)

NB: not all object classes are things (i.e. instance specific)

NB: not all object classes are things (i.e. instance specific)

Thanks, it's clear now.

and we probably (hopefully) never observed two trains on the same rail (or also in adjacent ones)

Do you mean scenes like this or this?

Nice find 🙂
But the points still hold true. It's very rare and therefore we turned instance separation off.

I only looked at few images, didn't search for long frankly. I'm writing a script to combine datasets and will probably make instances out of on-rails in MVD to make it compatible with other datasets. Thanks for making MVD publicly available.