yasinalm/gramme

We provide the code, pretrained models, and scripts to reproduce the experiments of the paper "Towards All-Weather Autonomous Driving". All code was implemented in Python using the deep learning framework PyTorch.

gramme's Issues

Hi， I hvae one question about the paper

In your paper, you said "Our novel multimodal geometric reconstruction algorithm and reciprocal training technique create a supervisory signal for the self-supervised neural network." I was wondering what is that technique and how to implement the reciprocal training technique. My understanding is that through this training technique, information can be exchanged between different modules, I don't know if this is correct, if it is correct, please tell me how to implement this technique, if not, please correct me.

Instructions for inference are incorrect

Thanks for sharing the code. I am trying to run an inference test for odometry trajectory using test_mono.py script with a pretrained Radiate dataset. However, instructions provided in the readme file are not correct.

Firstly, test_mono.py should specify in the arguments --pretrained-pose and not --pretrained-disp (currently in the readme):
python test_mono.py /path/to/dataset/ --dataset dataset_name --pretrained-disp /path/to/pretrained_model --results-dir /path/to/save/results # for monocular odometry

Secondly, I get the following error trying to run the script. I guess sequence parameter is missing however, it is not clear from description what sequence is and what it should contain:
python3 test_mono.py ./datasets/radiate_city_1/ --dataset radiate --pretrained-pose mono_posenet_checkpoint.pth.tar-20221020T225942Z-001/mono_posenet_checkpoint.pth.tar --results-dir ./results
=> creating model
=> using pre-trained weights for PoseResNet
Traceback (most recent call last):
File "test_mono.py", line 260, in
main()
File "/home/user/.local/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
return func(*args, **kwargs)
File "test_mono.py", line 128, in main
raise argparse.ArgumentError(
TypeError: init() missing 1 required positional argument: 'message'

Please can you update the instructions and specify correct arguments for running the inference tests.

research paprer not available

Hi,
i tried to download this research paper mentioned in this repo but it's not availabe. could you please share this paper.
secondly, results for radiate dataset only with stereo are good but others such as mono and lidar+camera results are bad. i tried to test using your provided trained model but results are still not good showing arbitrary and distorted trajectories.

Stereo only:

lidar + camera for radiate dataset pretrained model results:

1. mono2radarpred_notgt

2. monopred_nogt:

3. radarpred_nogt:

looking forward to your feedback.

Thanks

train.txt and val.txt

Thanks to the author for sharing the code. I want to know where train.txt val.txt is, or how it is divided