우수논문상, 단일 파노라마 입력의 실내 공간 레이아웃 복원 모델 경량화
졸업논문발표회, SoongSil University School of AI Convergence, 2021
We propose a method to lightweight the feature extraction network of HorizonNet by replacing ResNet and LSTM with MnasNet and GRU.
HorizonNet에 사용되는 특징 추출 네트워크를 최적화하여 경량화된 모델로 단일 파노라마 입력에서 실내 공간 복원 방법 제시.
Proposed Method
we propose a method to lightweight the feature extraction network of HorizonNet by replacing ResNet and LSTM with MnasNet and GRU.
- LSTM -> GRU
- ResNet50 -> MnasNet
Results
Quantitative Results The parameters used in the calculation have decreased by more than 1/2. At the same time, it was confirmed that there was no significant difference from the existing models in the performance measurement of 2D IOU, 3D IOU, MSE, Pixel error, and Corner error. MnasNet has a great influence on lightweighting, and GRU has an impact on accuracy.
Quantitative Results
The parameters used in the calculation have decreased by more than 1/2. At the same time, it was confirmed that there was no significant difference from the existing models in the performance measurement of 2D IOU, 3D IOU, MSE, Pixel error, and Corner error.
MnasNet has a great influence on lightweighting, and GRU has an impact on accuracy.
| Test cases | ResNet50-LSTM | MnasNet-LSTM | *Our MnasNet-GRU |
|---|---|---|---|
| #Parameter (Total) | 81,570,348 | 40,397,700 | 37,641,092 |
| #FLOPs | 71.83 | 59.19 | 58.48 |
| 2D IOU (%) | 87.07 | 84.16 | 85.07 |
| 3D IOU (%) | 84.53 | 80.80 | 81.89 |
| RMSE | 0.18 | 0.23 | 0.21 |
| Pixel error (%) | 2.04 | 2.57 | 2.70 |
| Corner error (%) | 0.65 | 0.80 | 0.84 |
Qualitative Results
Similar performance to existing models.
