any ideas how to overcome the below? Loaded weights from logs/7Scenes_heads_mapnet++_mapnet++_7Scenes/epoch_005.pth.tar Running mapnet++ on VAL data Traceback (most recent call last): File "eval.py", line 124, in vo_func=vo_func, no_duplicates=False, **kwargs) File "../dataset_loaders/composite.py", line 45, in init self.dset = SevenScenes(*args, real=self.real, **kwargs) File "../dataset_loaders/seven_scenes.py", line 52, in init with open(split_file, 'r') as f: IOError: [Errno 2] No such file or directory: '../data/deepslam_data/7Scenes/heads/TestSplit.txt'

Sir I run your code to Evaluate the pretrained Mapnet++ on Heads, Offices of 7 scenes and I get such jumbled output as compared to The output u show me in your research paper.

The output should not be just like in your research paper as below.?

https://github.com/NVlabs/geomapnet/blob/e18b75684ae61ca502e7df0cb99edf45770ac57b/scripts/process_robotcar_images.py#L38

The argument here seems should be scene=args.scene rather than sequence=args.scene, according to the dataset_loaders/robotcar.py https://github.com/NVlabs/geomapnet/blob/e18b75684ae61ca502e7df0cb99edf45770ac57b/dataset_loaders/robotcar.py#L19

Dear Mr. Samarth Brahmbhatt Thank you for your code! It is fantastic. However, I repeated three times, starting all over again (without resume checkpoint), training 0 to 100 epochs with mapnet, and using eval.py to test the trained model.

In theory, I train in strict accordance with the provided experimental parameters and compiling environment. The three pieces of training are completely independent and do not affect each other.
So the test results (for example, using eval.py to test the 100_epoch.pth.tar of the first, second, and third training respectively), the results should be almost the same. Unfortunately, the trained model (e.g. epoch_100.pth.tar) obtained in the second time's training gave an obviously larger error than that from the first-time's training. So do the third-time's results comparing with the second and the first time's ones.

I feel very confused. I'd like to ask for your opinion. Your suggestion will be very helpful to me. Thank you in advance!

Best Jialu

Hey, I noticed that the pose stats of 7-Scenes, namely the mean_t and std_t are simply set to zeros(3) and ones(3). But when I use the pose stats computed by myself (not equal to zeros(3) and ones(3)) for training and evaluation, the accuracy decreased a lot.

@samarth-robo Sir, I want to ask What is the difference between frame00xxx-pose.txtfile 4x4 pose matrix for each corresponding image and DSO: Direct sparse odometry txt filesof 12 values of poses for each corresponding image.??

Does your code needs to compare the results from both of them used as a ground truth.? why your code needed 2 poses files for the comparison.? what is the difference between these two files.??

Kindly guide me. I will be thankful to you.. Best regards

@samarth-robo, as you can see in this screenshot, The loss turn into negative sign.? what is it mean.? when I try to train MAPNET on my GPU Nvidia RTX2080ti with these parameters as below

n_epochs = 300 batch_size = 20 do_val = yes seed = 7 shuffle = yes num_workers = 5 snapshot = 50 val_freq = 50 max_grad_norm = 0

@samarth-robo Hello~ I have run this code with my own dataset, I choose Mapnet to be my model, hole settings are the same(learning rate, step, skip, etc..), but the result always converge to the same point, is this overfitting. I think the reason is

1. Learning rate still too big?
2. The parameter of skip is too small? (Because my robot move slow)
3. Data repeatability is too high (It means the image is too close but the poses are different because the path of my training data is close)

Following is the Ground Truth of my training data which do you think is the problem I really need your help thanks

Hi Samarth, I am implement your paper and I have a question for the initializations about Beta in mapnet.ini .I found that you set Beta = -3.0 in mapnet.ini ,and in your paper (Section 3.5. Implementation Details) you choose Beta = 0.0 instead.

Can I learn from you that which is the better initialization for Beta, and whether it has some reason to choose that initialization? Thanks, waiting for your reply.

@samarth-robo, Respected Sir, I have successfully installed your code and it works fine on 7-Scenes dataset,

Now I want to run your code on my own Images, I have extracted the frame00xxx-pose.txt for every corresponding images by using Visual-SFM, but here your code also needs the DSO: Direct sparse odometry of 12 values for the comparison, so can you please guide me how to extract the DSO from monocular images having 4x4 poses for each frame too..??

Or can you please guide me format of DSO file values,

Is these values are also Rotation and Translation.? and what is the format of these files in that text file named "dso_poses/seq-xx.txt".?? like 12 values shows what.? which values are for which parameters for camera.?

waiting for your kind response, Best regards.

@samarth-robo Respected Sir, As this code run on ur dataset of "7-Scenes" now if we want to run this code on our custom images then How would this code will able to evaluate and predict the results.?

as in Dataset of "7-Scenes" It needs two things as an input. 1 is "RGB image" 2 is "frame-000001.pose.txt" files having 4x4 pose matrix.

so in our own custom images, Can you please guide me how to get this 4x4 pose matrix "frame-000001.pose.txt" from every RGB images.?

I will be thankful to you and waiting for your kind response.

Regards

Hello, excuse me. For some special reason I have to use python3.6+pytorch1.6 for inference on mapnet, but to my surprise the results in this case are worse than inference with the environment in the environment.yml file Many, what is the reason for this, if you can answer, I will be very grateful.

Hi, thanks for sharing such a nice work. I have some trouble downloading the 7-Scenes dataset. The download links in https://www.microsoft.com/en-us/research/project/rgb-d-dataset-7-scenes/ does not work. Can you share your data? Thanks.

Hi,

Thanks for your great work. May I know what kind of algorithm is used in pose interpolation? def interpolate_poses(pose_timestamps, abs_poses, requested_timestamps, origin_timestamp):

Hi there, @samarth-robo. Thanks for your solid work and your well-structured code, I could produce even better results than the numbers in the original paper by loading pre-trained model weights.

However, when I retrained on the RobotCar dataset and loop scene by using the provided script and config file (from latest version): python train.py --dataset RobotCar --scene loop --config_file configs/mapnet.ini --model mapnet --device 1 --learn_beta --learn_gamma

The results become weird and errors are much larger than I expected.

It's worth noting that, I executed the script on an 8 * NVIDIA RTX 2080ti node: When I used pytorch-0.4.1, which is specified in your environment.yaml, there was an error detected by cuda: "THCudaCheck FAIL file=/pytorch/aten/src/THC/THCGeneral.cpp line=663 error=11 : invalid argument" Thus I ran the script both in pytorch 0.4.1 and 1.0.1 environment. However, both errors are very large.

Besides, I also noticed that the preprocessed images have some over-exposure cases (some are almost all white and barely has information), is it normal?

like 1403774724292807.png, ... 1403774724917727.png at the beginning of 2014-06-26-09-24-58 and other sequences.

Hey, I know that you have addressed this a number of times on this forum but I have a few doubts of my own. First, I would like to know what exactly I would have to introduce when I implement the algorithm with my own dataset. For example, the seven scenes dataset contains three types of data -RGB images -Depth images -Pose text file How do you obtain said text file for frames of a custom dataset. Beyond that, what else would I have to change in the repo if I want to run it on my personal dataset While moving through the issues, I noticed a lot of queries about the DSO files and xxxx_vo_stats.pkl files. Any help I could get would go a long way. Thanks in advance