ddavid
commented
4 years ago Hey there Lukas, firstly the quality of the labels is great on the whole, good job! It looks like your team had a lot of fun while creating the first samples.
There's a few things before merging though:
- Consistency within your dataset:
Although both images above are very similar, the blue cones labeled in the back swap. Why not label both in each image?
For the skid-pad images you've labeled even the smallest of cones, but for the sequence coming into and back from Acceleration a lot of cones that look similar are left out.
I'd consider keeping the latter out of your dataset, if you're still above the 600 image contribution threshold without them.
Along with 817 and 818, I'd expect you to label the cones being carried from how you labeled the rest.
-
Wrong labels, the bounding box is put incorrectly:
-
Wrong Class:
Another point is the small orange/red cones from the FSS and FSG footage. I'm afraid those would need to be labeled as otherwise it would be hard to use the dataset together with others. Not having those classes is possible, as long as they don't appear in the dataset's images. The blue-cones in FSS are different enough to be disregarded, since they don't have the white stripe and the hue is quite different.
Lastly, there were a few labels that were mal-formatted. Maybe something went wrong when parsing them? mal-formatted.txt
lukaskoenen
Initial FaSTDa-Racing lables and Documentation for Review.