Eval on a custom point cloud

[KevinCain]

I'm attempting to process one of the provided point clouds into a mesh via dgnn as a sanity check before trying my own numpy output. The features seem fine (see log below), however run.py gives an error (also shown below). My steps follow:

• Set data path in configs/custom.yaml, e.g.: data: /home/kevin/dgnn/data/sample
• Extract features with precompiled feat build provided:
  ./utils/feat -w /home/kevin/dgnn/data/sample/ -i anchor_0 -s npz
• Attempt to create mesh:
  python run.py -i -c configs/custom.yaml
  The error returned is:

Traceback (most recent call last):
File "/home/kevin/anaconda3/envs/dgnn/lib/python3.9/site-packages/munch/__init__.py", line 103, in __getattr__
return object.__getattribute__(self, k)
AttributeError: 'Munch' object has no attribute 'metrics'

The feature extraction log follows:

-----FEATURE EXTRACTION-----

Working dir set to:
-/home/kevin/dgnn/data/sample/

Read NPZ points...
-from anchor_0
-3110 points read
-with sensor
-with normal

Export points...
-to anchor_0.ply
-3110 points

Delaunay triangulation...
-3110 input points
-3110 output points
-39342 finite facets
-19520(+604) cells
-done in 0s

d parameter of Labatu set to -1

Consider turning on scaling with --sn, if your learning data is not yet scaledto a unit cube!

Start learning ray tracing...
-Trace 1 ray(s) to every point
-Ray tracing done in 0s

Export graph...
-to gt/anchor_0_X.npz
-Exported 20124 nodes and 80496 edges
-in 1s

Export 3DT...
-to gt/anchor_0_3dt.npz
-3110 vertices
-39342 facets
-19520 tetrahedra

Export interface...
-to anchor_0_initial.ply
-Before pruning (points, facets): (3110, 39342)
-After pruning (points, facets): (304, 604)
-done in 0s

-----FEATURE EXTRACTION FINISHED in 1s -----

[KevinCain]

A quick note: as per dgnn's environment requirements I have munch (2.5.0):
/home/kevin/anaconda3/envs/dgnn/lib/python3.9/site-packages
Python finds module munch as /home/kevin/anaconda3/envs/dgnn/lib/python3.9/site-packages/munch/__init__.py

[KevinCain]

I converted the supplied .ply data to .npz rather than using the .npz files above, in case the normal and viewpoint data was not present:

cd ~/dgnn
python ./processing/reconbench/ply2npz.py --user_dir=/home/kevin/dgnn/data

This outputs files as expected, and the feature extraction from the output looks sane, similar to what I logged above.

However, I get the same error noted above when calling run.py for evaluation.

[raphaelsulzer]

Hi, the custom.yaml file was not really up-to-date.

I think you are simply missing this line in it.

[KevinCain]

Thanks @raphaelsulzer, I did a hard reset via git and pulled the latest updates, which included the line you noted. I'm able to proceed, but quickly found that I do not understand how to set the inference list in the case of single point cloud.

As above in this thread, I converted a single point cloud via ply2npz, which I placed in the ‘0’ folder expected in ground truth data, similar to the five model group (0..4) from 'reconbench'. I see run.py still expects a batch of five point models.

Consider this invocation: python run.py -i -c configs/custom.yaml, where the .yaml inference settings include:

inference:
  dataset: sample
  classes: anchor

Based on the code, it looks like the above expects the following directory structure:

...data/
├─ sample/
│ ├─ anchor_0.ply
│ ├─ anchor_0.npz
│ ├─ anchor_0_initial.ply
│ ├─ gt/
│ ├─ ├─0/
│ ├─ ├─├─ anchor_0_labels.npz
│ ├─ ├─├─ anchor_0_* ...

Do I have the .yaml or folder structure wrong?

How can I instruct dgnn to process the single file group above? For training and validation, the analogue seems to be scan_confs: x.

[raphaelsulzer]

You can structure your dataset any way you want as long as this dictionary is correctly filled. I recommend you to use a python debugger and see if all the paths are correct.

Also your example should probably look like this:

inference:
  dataset: sample
  classes: ["anchor"]

[KevinCain]

Thanks once again for a fast, and kind, response. I had mistakenly switched label data on; once I ran with has_label: 0 I was able to process a single input file as expected. I added a block in dataset.py to handle inference from my dataset 'sample' shown below:

READ CONFIG FROM  /home/kevin/dgnn/configs/custom.yaml
SAVE CONFIG TO  /home/kevin/dgnn/data/models/kf96/config.yaml
dataset: sample
class:  ['anchor']
Point clouds per class:  0
path:  /home/kevin/dgnn/data/sample
filename:  anchor_0
scan_conf:  0
gtfile:  gt/0/anchor_0

Thanks again @raphaelsulzer for your help!