search

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
338 stars 97 forks source link

The detection performance on TLESS dataset is very poor. What's wrong? #45

Closed AlbertLZG closed 4 years ago

AlbertLZG commented 4 years ago

Hi, I want to reproduce results in paper. So I trained a retinanet following the instruction:

  1. Generate a training dataset by running 'python detection_utils/generate_sixd_train.py'
  2. Train a retinanet according to ' https://github.com/fizyr/keras-retinanet' However, I only get a mean ap 18% on TLESS dataset at the threshold IOU=0.5 So, What could be the problem? I would be grateful if you could provide detailed configuration of training retinanet.
MartinSmeyer commented 4 years ago

Hmm, difficult to tell. How is the performance on the training set? Did you change from training to test mode (batch norm)? Did you fine-tune from a checkpoint?

MartinSmeyer commented 4 years ago

Sorry, currently I don't have time to write a step-by-step manual to train a detector. Please refer to the explanations in the paper (latest version) and https://github.com/fizyr/keras-retinanet

Closing for now.