On Human Motion Prediction Using Bidirectional Encoder Representations from Transformers
碩士 === 國立臺灣科技大學 === 電子工程系 === 107 === Pose prediction found applications in a variety of areas. However, current methods adopting recurrent neural networks suffer from error accumulation in the training stage. Furthermore, encoder-decoder architecture in general fails to predict continuous poses bet...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | en_US |
| Published: |
2019
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/du2jv4 |
| Summary: | 碩士 === 國立臺灣科技大學 === 電子工程系 === 107 === Pose prediction found applications in a variety of areas.
However, current methods adopting recurrent neural networks suffer from error accumulation in the training stage. Furthermore, encoder-decoder architecture in general fails to predict continuous poses between the end of the encoder input and the beginning of the decoder output.
Benefiting from the recent successes of the attention mechanism, in the thesis, we propose a novel method which combined the transformer encoder architecture and universal transformer.
The new architecture is free of error accumulation because this architecture processes data parallelly and the weight of updating for each position is equal. Moreover, the proposed attention map helps attention mechanism to refrain the predicted poses from discontinuity.
We also apply adaptive computation time algorithm to optimize the iteration numbers of performing an attention mechanism.
The mean absolute loss is considered to handle human motion prediction problem in the training process on the Human3.6M dataset.
Simulations show that the proposed method outperforms the main state-of-the-art approaches.
|
|---|