This file is in TODO
Related works:
- Optimize Vision Transformer Architecture via Efficient Attention Modules: A Study on the Monocular Depth Estimation Task
- EfficientViT: Memory Efficient Vision Transformer with Cascaded Group Attention
This work aimed to adopt EfficientVit architecture, on METER depth estimation models, to analyze performance and seek for improvements.
Original METER models are:
- METER (encoder + decoder)
- MetaMETER
- PyraMETER
Variations studied in this work:
- EfficientVit + decoder METER
- Encoder (MobileVit + EfficientVit) + decoder METER
- MetaEfficientVit + decoder METER
- PyraEfficientVit + decoder METER (TODO)
| Model | RMSE | MAE | Abs rel | δ1 | δ2 | δ3 | Parameters number |
|---|---|---|---|---|---|---|---|
| METER | 48.440 | 36.270 | 0.143 | 0.819 | 0.959 | 0.988 | 3,295,168 |
| MetaMETER | 50.758 | 37.860 | 0.150 | 0.802 | 0.950 | 0.988 | 3,220,864 |
| PyraMETER | 49.159 | 36.690 | 0.145 | 0.809 | 0.957 | 0.987 | 5,527,168 |
| Model | RMSE | MAE | Abs rel | δ1 | δ2 | δ3 | Parameters number |
|---|---|---|---|---|---|---|---|
| EfficientVit | 58.878 | 44.828 | 0.172 | 0.734 | 0.933 | 0.982 | 2,627,668 |
| MobileEfficientVit | 47.525 | 35.344 | 0.138 | 0.825 | 0.962 | 0.989 | 3,288,415 |
| MetaEfficientVit | 51.162 | 38.462 | 0.149 | 0.793 | 0.950 | 0.987 | 3,115,696 |
| PyraEfficientVit | 47.710 | 35.718 | 0.140 | 0.818 | 0.963 | 0.989 | 3,309,115 |
Note: GPU is NVIDIA RTX GeForce 4090
| Model | Average CPU (ms) | P90 CPU (ms) | Average GPU (ms) | P90 GPU (ms) |
|---|---|---|---|---|
| METER | 17.879 | 18.679 | 4.944 | 4.186 |
| MetaMeter | 15.370 | 15.770 | 3.625 | 3.792 |
| PyraMETER | 15.203 | 16.401 | 4.465 | 4.573 |
| Model | Average CPU (ms) | P90 CPU (ms) | Average GPU (ms) | P90 GPU (ms) |
|---|---|---|---|---|
| EfficientVit | 13.305 | 13.941 | 8.367 | 8.430 |
| MobileEfficientVit | 15.663 | 16.304 | 7.031 | 7.185 |
| MetaEfficientVit | 12.441 | 12.821 | 2.913 | 3.114 |
| PyraEfficientVit | 11.081 | 14.125 | 8.759 | 9.735 |