Skydivers are not recognized as humans at all from back view.

#3
by Rollerblade128 - opened

I just checked whether the model can segment a skydiver from a back view. It was not able to identify any segments although the two legs, two hands, were clearly seen.

AI at Meta org

@Rollerblade128 is it possible to share the image you tested on? Thanks.

@rawalkhirodkar of course I can share with you. I found also scuba diving pictures where the model suffers.
Discover_Scuba_Diving_--_St._Croix,_US_Virgin_Islands.jpg

O.1.4_00108004.jpg

AI at Meta org

@Rollerblade128 these images are challening for our segmentation network which expects primarily single human image with good visibility.

rawalkhirodkar changed discussion status to closed

@rawalkhirodkar Do you think it makes sense to finetune the model by couple of hundreds images?

AI at Meta org

@Rollerblade128 yes, i think this will definitely help. You have the data ready, please checkout the github repo for finetuning on custom segmentation labels.

Thanks @rawalkhirodkar .are you aware of any web ui tool which can help to make manual annotations less painfull. My datasource is a collection of videos. Spliting them to frames by opencv is trivial, but the annotation is jogh resource activity. Every egficiency improvment could help a lot

Sign up or log in to comment