search

dineshreddy91 / carfusion_to_coco

This repository is to setup the Carfusion Dataset for training keypoint detector for cars
Other
30 stars 11 forks source link

Key points information #3

Closed XiongZhongxia closed 3 years ago

XiongZhongxia commented 4 years ago

Hi, @dineshreddy91. Thanks for your contribution!

I only find 12 rather than 14 points according to the sample images. Another problem is that which parts (e.g. left headlight， right rear light...) of a vehicle are considered as key points in Carfusion. Could you provide more information for annotated key points？

XiongZhongxia commented 4 years ago

Never mind. I found key point names in Occlusion_Net/lib/data_loader/datasets/keypoint.py. 

NAMES = [
    'Right_Front_wheel',
'Left_Front_wheel',
'Right_Back_wheel',
'Left_Back_wheel',
'Right_Front_HeadLight',
'Left_Front_HeadLight',
'Right_Back_HeadLight',
'Left_Back_HeadLight',
'Exhaust',
'Right_Front_Top',
'Left_Front_Top',
'Right_Back_Top',
'Left_Back_Top',
    'Center'
]

I have also checked several annotation files in CarFusion, I'm pretty sure that 'Exhaust' and 'Center' points are not included in GT files. So actually there are annotations of only 12 points. But why the channel dimension of kp_logits in Occlusion_Net/.../graph_head.py is 14？？

dineshreddy91 commented 4 years ago

We labelled 14 points with the exhaust and the center. but they were hard to annotate and were giving ambiguous results. We ignore both of these points in most of our evaluations and visualizations and concentrate on the 12 points which are consistent and used across all of our frameworks. in case of Occlusion_Net/.../graph_head.py we use 14 as a buffer and generally ignore the output for exhaust and the center.