search

microsoft / singleshotpose

This research project implements a real-time object detection and pose estimation method as described in the paper, Tekin et al. "Real-Time Seamless Single Shot 6D Object Pose Prediction", CVPR 2018. (https://arxiv.org/abs/1711.08848).
MIT License
727 stars 214 forks source link

I would like to ask, when pre-training initialization weights, you need to pay attention to the big angle error, as shown below #110

Open yunxijun opened 5 years ago

yunxijun commented 5 years ago

Screenshot from 2019-08-25 14-53-02

banannakn commented 5 years ago

I got the same problem, my angle error is stable or slightly increasing, everything but descending.. Screenshot (15)

btekin commented 5 years ago

During pre-training, we do not train for confidence prediction but only for positional prediction. This is to ensure that when we start training for confidence estimation, we have reasonable positional estimates. Early during training, positional accuracy would be not so high, the confidence values computed using Eq. 1 during training would be almost always low and therefore training for confidence early during training would not make sense. Since we do not train for confidence value during pre-training, the accuracy of the highest confidence prediction would not make much sense at pre-training. However, you should definitely take into account the translation and angle errors during the training after the initial "pre-training" phase.