Pedestron
2024-01-05 10:45:22
The Pedestron repo of Irtiza Hasan is used to train a Cascade R-CNN with MobileNetV2 backbone (smaller than the HRNet used by them) on manipulated training data. The training data is manipulated in a way that ambiguous samples are excluded. In my bachelors thesis, I am investigating the impact of ambiguous data on the training but also on the evaluation and LAMR. I am doing all of the experiments on the validation data but I would like to have a comparable LAMR value on the test dataset. On the non-manipulated validation data, the model achieves the following LAMRs: [0.114, 0.2, 0.499, 0.274]. It was only trained for 50 epochs (original Pedestron is something like 147 epochs) and no external data is used for pretraining.
| Method description: | The Pedestron repo of Irtiza Hasan is used to train a Cascade R-CNN with MobileNetV2 backbone (smaller than the HRNet used by them) on manipulated training data. The training data is manipulated in a way that ambiguous samples are excluded. In my bachelors thesis, I am investigating the impact of ambiguous data on the training but also on the evaluation and LAMR. I am doing all of the experiments on the validation data but I would like to have a comparable LAMR value on the test dataset. On the non-manipulated validation data, the model achieves the following LAMRs: [0.114, 0.2, 0.499, 0.274]. It was only trained for 50 epochs (original Pedestron is something like 147 epochs) and no external data is used for pretraining. |
| Computation Environment: | one RTX4090 |
| Computation Time: | 341724.0 |