lzccccc / SMOKE

SMOKE: Single-Stage Monocular 3D Object Detection via Keypoint Estimation
MIT License
696 stars 177 forks source link

can you explain the targets attributes with trans_mat means when inference ? #61

Closed SmallMunich closed 2 years ago

SmallMunich commented 2 years ago

Hi, great works. I debug this code when I confused about the targets[i].get_field("trans_mat"). When to inference, the targets[i].get_field("trans_mat") represent what?

poledna commented 2 years ago

Hi, have you ever figured trans_mat out ? What does it means other than

    trans_mat = get_transfrom_matrix(
        center_size, 
        [cfg.INPUT.WIDTH_TRAIN // cfg.MODEL.BACKBONE.DOWN_RATIO,
         cfg.INPUT.HEIGHT_TRAIN // cfg.MODEL.BACKBONE.DOWN_RATIO]
    )

Im really confused because my pixel bbox prediction is correct but the placement of the keypoint is way off, i believe it might be this trans_mat. Thanks