search

lzccccc / SMOKE

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

Result not very good on testing #15

Closed lucasjinreal closed 4 years ago

lucasjinreal commented 4 years ago

Some sample image visualized

image image image image image

miss detection is seriously problem.

Florijan-Iljazovic commented 4 years ago

Hey man I'm just wondering did you use some existing repo for these nicely visualized boxes ? Or is it your custom work. Thanks! I like that the front side is shaded, and I'm lazy to do it myself.

beedotkiran commented 3 years ago

Hello @jinfagang,

We were using models trained on the train-set to predict on KITTI test set, and had poor detection results at the center of the image on nearby/close objects (most cases), as in your case.

We also tried training and predicting on the Lyft dataset in KITTI format. We observed the same effect missing predictions on nearby large objects.

Could you please share how you have resolved this problem ?

Thank you!

nikhil-nakhate commented 3 years ago

Hi @beedotkiran , Did you have a different set of classes in the Lyft dataset. I am not able to train on a dataset with different classes. Any help would be appreciated.

Best Regards, Nikhil