JOIN
Get Time
forums   
Search | Watch Thread  |  My Post History  |  My Watches  |  User Settings
View: Flat (newest first)  | Threaded  | Tree
Previous Thread  |  Next Thread
Data evaluation | Reply
1 - Some .json files in debris_or_rubble_154 have a negative bounding box origin... is that a small typo ? Can I fix this problem setting the negative value to zero ?
2 - What does the visible field in the bounding box mean ? Initially I though it would be false if the image was not visible but I found some cases that I could clearly see the bounding box image.
Re: Data evaluation (response to post by thomio.watanabe) | Reply
Response from the organizers:
1. Yes, setting the negative value to zero seems like an acceptable fix.
2. The visible tag is meant to indicate if the feature of interest is visible in a specific image. For example, the flooded road category might have several temporal views of a road, but the visible tag should only be true for the images where the road is actually in a flooded state. Similarly, a construction site sequence might have temporal views of the site after construction is finished, but in those views where the bounding box no longer contains a construction site the visible tag should be false.
Re: Data evaluation (response to post by walrus71) | Reply
Thanks walrus, this seems to be very important.
I think you just answered this but to make it clear...
If there is a test image of a road that is not flooded, should it be classified as flooded_road or false_detection ?
I was wondering what should be considered false_detection, but it seems that all bounding boxes where the visible tag is false should be considered false_detection, right ?
Re: Data evaluation (response to post by thomio.watanabe) | Reply
You can assume that if a bounding box is classified as flooded_road then at least one of the temporal views of that scene will indeed show a road that is flooded. Like discussed above it may not be the case for all images that feature the same bounding box.
For false detections all temporal views that feature the same bounding box will show a scene that can not be classified as any one of the known categories. E.g. if all views show only woods or desert then this is surely a false detection.
RSS