Missing dataset_info.yml (For YCB-Video)

[lachwot]

Hi! When running predict.py, I get the following error if I don't specify a path for train_data_path:

Traceback (most recent call last):
  File "predict_original.py", line 632, in <module>
    with open(dataset_info_path,'r') as ff:
FileNotFoundError: [Errno 2] No such file or directory: 'None/../dataset_info.yml'

This makes sense, given that None as a path yields nothing. Setting train_data_path to where I have set up my dataset however, gives me the same error as dataset_info.yml cannot be found in my YCB-Video dataset folder.

I can confirm that I have downloaded the dataset correctly since my teammate and myself have the same files and folders within the YCB_Video_Dataset folder.

Should I have a dataset_info.yml file somewhere? If not, how should it be set out? Is there a file that I can download? What should my YCB_Video_Dataset folder look like? Here is my current directory within YCB_Video_Dataset:

cameras  data_organized  image_sets  LICENSE  pairs  README
data     data_syn        keyframes   models   poses

I should also mention that my dataset folder is separate from this repository's folder due to me needing it to test other 6D pose estimator networks. Looking at the code in more detail suggests that dataset_info.yml is required for predictions as it contains camera parameters, image resolutions, bounding box info, etc.

Cheers, Lachlan

[wenbowen123]

There is a similar issue mentioned here https://github.com/wenbowen123/iros20-6d-pose-tracking/issues/10. It's likely that the arguments in predict.sh is not passed correctly in the predict.py. I will correct that when I find a chance. For quick fix, you can directly change the default params.

[lachwot]

Cheers for the reply!

I'm checking those out at the moment, and I can see that predict.sh has training paths in it. Looking at Issue #10, I can see that your synthetic data is used as the training path (of which each object does indeed include a dataset_info.yml file). I've updated mine to their correct respective paths.

Does there exist a dataset_info.yml file for the "raw" YCB-Video Dataset?

For some context I should also mention that I only wish to infer on some of the YCB-Video dataset images (I'm aiming to rig the network to an RGB-D Kinect 2 camera feed with a ROS node), so I'm hoping to not have to download each object's large zipped synthetic data file. This would also mean I wish to track multiple objects at once (multiple objects inferred for each frame).

Other than that, I think I have the file running, at least for now :P

[lachwot]

I might also take the chance to ask now, is it possible to have a zipped folder of the dataset_info.yml and textured.ply files for each of the YCB-Video objects uploaded somewhere? Just so I can mess around with those.

Cheers

[wenbowen123]

good to see you are able to run. If you want to test on your own data from Kinect, just a reminder that the trained data are rendered using different instrinsic params from your Kinect's. Not sure if it still works. The recommended way is to re-generate some synthetic data using the same intrinsic params as your own camera. This is how we did for YCBInEOAT dataset.

Unfortunately I dont have such collections. You may have to download the zip folders of your interested objects.

[lachwot]

Does this mean I'll have to retrain the network using data generated with my camera's own intrinsic parameters? Or can I get away with knowing my camera's parameters (I've calibrated it), and use your weights on the "raw" YCB-Video dataset/ Kinect 2 feed? The original YCB-Video dataset almost certainty would have been filmed on a camera with different parameters to our cameras. I know that the dataset_info.yml files for each object have sections for the camera's intrinsic parameters though I don't know if these are exclusively for training. I presume that I can just update those camera parameters to what my Kinect2's parameters are, and then feed in my camera feed to the network.

Here is the current dataset_info.yml file for now, with your original parameters.

boundingbox: 10 
camera:
  centerX: 312.9869  
  centerY: 241.3109  
  focalX: 1066.778
  focalY: 1067.487
  height: 480
  width: 640
distribution: gauss
models:
- 0: null
  model_path: /home/mu00185683/iros20-6d-pose-tracking/models/052_extra_large_clamp/textured.ply
object_width: 238.6939428307838
resolution: 176
train_samples: 200000
val_samples: 2000

[wenbowen123]

To get the better performance, it's better to regenerate the synthetic data using your latest camera params and retrain on it. But I haven't tried directly updating the camera params. It might work if your params doesn't differ too much, otherwise the images captured by two very different cameras may look drastically different. Since I haven't tried in practice, these are my guessings.

[lachwot]

Ah, I see. I was hoping to avoid having to generate my own data, so I'll stick to the dataset_info.yml files that you have provided as a part of your data for now. Thanks for all the help!

[wenbowen123]

You can check YCB-Video and YCBInEOAT dataset and choose whichever's intrinsics that matches closer to your own camera.

[jingweirobot]

@lachwot Hi Dr.Wen. So I also have the similar question. How could you generate the dataset_info.yml files. In the raw ycb-video dataset, there are not such files. Thanks in advance.