roym899 / pose_and_shape_evaluation Goto Github PK

Currently a single view per dataset is autogenerated with prediction on top of point cloud. Better to generate three separate figures: prediction, ground-truth, and masked RGB-D, as shown below. Camera view has to be adjusted to automatically focus object of interest.

Possibly auto crop like this

There are some issues with the current metrics implementation.

Some thoughts:

• single estimate metrics / dataset metrics
• dataset metrics are typically defined by a single-estimate metric and aggregation function
• some metrics are composed of other metrics
• for some metrics bigger is better, for other smaller is better
• support swiping single-estimate metric parameters (like a threshold) and automatic plotting
• symmetry handling varies per-metric

Currently most of this is supported for the standard metrics, but it's quite messy.

Hi. Thank you for releasing such a wonderful work~
I'm trying to evaluate pretrained models such as CR-Net on the redwood dataset, and I have configured the dataset as required. However, when I run the evaluate code there is an error regarding the absence of instance masks? Do I need any additional processing for the RedWood dataset?
I'm looking forward to your reply!? looking forward to your reply.

Some of the included methods use jit compiled modules. This will cause slow initialization without any user feedback at the moment. It might also fail ungracefully if nvcc is not available.

Many methods use a get_bbox function which is very similar (probably exactly the same). It would be good to create a module with common functions used by many of the methods. Right now the image width and height is hardcoded for most methods, which should be changed.

Hello, thank you for your valuable work! I'm planning to test my own network using this toolbox, but when I install it, I faced the following problem:

ERROR: Could not find a version that satisfies the requirement yoco (from versions: none)
ERROR: No matching distribution found for yoco

I cannot find solutions from the internet, how to fix this? Is it caused by the version mismatch of the some package?

Hi, thanks for your excellent work! With the help of your tool, I have done most of the performance evaluation of the model on the REAL dataset.
But when I do the performance evaluation on the RedWood dataset, the obtained metrics look very different from those in the paper (take dpdn as an example, see the snapshot below).

I visualized the input of the dataset and found that the angle of the generated mask is not consistent with the angle in the RGB image,. I am not sure if this is the reason for this problem. Could you please tell me how to solve this problem?

Currently Python 3.6 support is limited due to slow KDTree implementation of scipy <1.5 (see 1.5 changelog). Possible fix might be to use scikit.neighbors module instead.

See #1 for initial discussion of this issue.