search

libornovax / master_thesis_code

Code for my master thesis: Vehicle Detection and Pose Estimation for Autonomous Driving
MIT License
186 stars 69 forks source link

NMS!! #31

Closed libornovax closed 7 years ago

libornovax commented 7 years ago

Merging is too slow and I don't think that the classical NMS can really use the potential of the detector...

libornovax commented 7 years ago

Updated confidence - try if we could use some information from the coordinates to "update" the confidence of the bounding boxes

libornovax commented 7 years ago

Probably not going to help as there does not seem to be additional information:

P(bb|car) vs P(bb|not car) hist_wh_bb_g_car_notcar_acc_x4

P(car|bb) hist_wh_car_g_bb_acc_x4

P(not car|bb) hist_wh_notcar_g_bb_acc_x4

libornovax commented 7 years ago

Does not seem to change anything in performance. If we used extended confidence conf = P(car|conf) + P(car|bb) it is still the same, maybe a little worse:

Normal confidence: pr_jura_test_short_fix

Extended confidence: pr_jura_test_short_fix

ansabsheikh9 commented 6 years ago

Hello Mr. Novak, great work, Could you tell me how are you performing NMS on oriented 3D boxes? BR, Ansa

libornovax commented 6 years ago

Hi, I chose perhaps the simplest solution - project the 3D boxes on the ground plane, create axis-aligned 2D bounding boxes on the ground plane (as an approximation of the oriented ones) and then perform classical NMS on them. Hope that helps.

ansabsheikh9 commented 6 years ago

Thank you for your reply. 👍 Best Regards, Ansab