retinanet accuracy about pedestron HOT 11 CLOSED

Yes, the model we published for RetinaNet is pre-trained on CrowdHuman. You can read more about it in our pre-print version. Furthermore, #3.

from pedestron.

@hasanirtiza

Thanks for your reply.
So I think the conclusion is RetinaNet is not powerful enough.
for face detection, retinanet can get very good accuracy, I don't know why it is not good in pedestrian detection.

Thanks
Meixitu

from pedestron.

We have some better numbers with the RetinaNet for pedestrian detection. We plan to release the model in future hopefully soon. Of course do not expect it to be up there with Cascade etc.

from pedestron.

One more comment, for the face detection, multi scale testing plays a very pivotal role for better accuracy.

from pedestron.

for your current published score, do you just use the original image size?

from pedestron.

Yes, most if not all models are evaluated using the original image size.

from pedestron.

Thank you very much

Thanks
Meixitu

from pedestron.

@hasanirtiza

You said you can get better performance for retinaNet, what did you modified? Could you give me some clue?

Thanks
Meixitu

from pedestron.

We will share it soon (keep checking the repo) , since we are still in the process training, as of now unfortunately we cannot share more details.

from pedestron.

@meixitu Checkout the README.md. A new model of RetinaNet has been added.

from pedestron.

@hasanirtiza ,

Got it.

Thanks
Meixitu

from pedestron.