search

NVlabs / planercnn

PlaneRCNN detects and reconstructs piece-wise planar surfaces from a single RGB image
Other
559 stars 125 forks source link

Are RGB images and annotations really aligned? #51

Open EryiXie opened 4 years ago

EryiXie commented 4 years ago

Hello, thanks for this excellent work. However, I am quiet confused about the rgb frames from ScanNet v2 and the plane annotations you offered.

For Example, scene0000_01 frame 33th: Annotations looks like: 33

while the rgb image looks like: 33

It seems the annotation is not aligned with the image. And I can not find any function in all the scripts inside ./dataset folder that solve the align problem. And the item pair from dataset and dataset_loader seems also doesn't align.

Did I misunderstand something?

ajithvcoder commented 4 years ago

@EryiXie could u tell the cuda version in which u got this output . In which cuda version did u install nms and roi

EryiXie commented 4 years ago

@EryiXie could u tell the cuda version in which u got this output . In which cuda version did u install nms and roi

Hi, this is not the output of PlaneRCNN. The "blue-green" mask map is actually the ground truth annotations from ScanNet. Same is the RGB image. As you can see, the ground truth does not really match the RGB image. That's the part confuses me a lot.

ajithvcoder commented 4 years ago

okay got it

PeterFWS commented 4 years ago

Clearly the ground truth provided by PlaneRCNN has large errors, which is reported from a ECCV 2020 paper "Learning Pairwise Inter-Plane Relations for Piecewise Planar Reconstruction". In this case, consider the experiment results from planercnn not reliable.

Karthikaddagadderamesh commented 3 years ago

@art-programmer When I generated from the annotation using the dat_prep/parser.py they are aligned with the RGB images. May be you can use that.