sem_seg learning rate clip typo

[themmes]

Dear Charles,

In sem_seg/train.py probably a typo which eliminates your learning rate clipping, which you so emphasize.

learing_rate should be learning_rate?

def get_learning_rate(batch):
--
def get_learning_rate(batch):
    learning_rate = tf.train.exponential_decay(
                        BASE_LEARNING_RATE,  # Base learning rate.
                        batch * BATCH_SIZE,  # Current index into the dataset.
                        DECAY_STEP,          # Decay step.
                        DECAY_RATE,          # Decay rate.
                        staircase=True)
    learing_rate = tf.maximum(learning_rate, 0.00001) # CLIP THE LEARNING RATE!!
    return learning_rate

[charlesq34]

Thanks for spotting that! I've fixed it.

[themmes]

You're welcome, did you also get a chance to look at the email I send you (charlesq34@gmail.com)?

Oh and btw, the typo is in all train.py, so also sem_seg/train.py

[charlesq34]

oops, thanks again.

What's the title of the email? I cannot find any..

[themmes]

PointNet++ multi-scale (2questions, 2minread) Sent on 6th of July to charlesq34@gmail.com

Maybe interesting to sidenote; I am currently preparing the semantic3d.net datasets to perform Semantic Segmentation using your model.

[themmes]

Odd, I have send you a new email. Let me know if it does not reach you.

[belevtsoff]

@themmes how did it go with semantic3d dataset? I'm really curious how well PointNet++ performs on the large scale outdoor data.

[themmes]

@belevtsoff Thanks for asking, I discussed this with Maxim Tatarchenko (@mtatarchenko, PhD from Freiburg). I discontinued the research on Semantic3D, mainly because I received another dataset (MLS, sadly not public) where I could exploit reference to the scanner. My train of thought here was that it could provide spatial reference in a similar way the PointNet on indoor exploits relative position in the room. I believe Maxim also discontinued the research on Semantic3D after getting results of about 17% accuracy.

[belevtsoff]

Oh, I see, thanks! Did you and Maxim do experiments with vanilla PointNet or pointnet++ as well? Seems to me like it could have a considerable improvement for large datasets.

[tatarchm]

Hi,

A colleague of mine and I only tried the vanilla PointNet. PointNet++ should definitely produce better results.

Best, Maxim

[belevtsoff]

Cheers guys. I'll let you know If I'll end up trying PointNet++ on some outdoor or aerial data.

[maximiliangoettgens]

@belevtsoff did you already perform the tests on outdoor data? I am currently working on a project on outdoor data and would be very interested in an exchange

[themmes]

@belevtsoff Yes, please keep us posted. I ended up moving to proprietary data of a highway scene with PointNet (achieved about 50% MIOU for four classes excl. background class) as the PointNet++ code was not yet released at that time.

@maximiliangoettgens I believe @tatarchm (or his colleague) results of PointNet on the Semantic3D dataset were about 20% MIOU for point-wise classification (or semantic segmentation / scene segmentation)

[belevtsoff]

@maximiliangoettgens @themmes Hey guys, just noticed the messages. Unfortunately, I had to drop out of the project, so I never ended up trying PointNet++ on the outdoor data. I'll forward this conversation to the guys from the project, maybe they've tried something by now.

[bw4sz]

bump on whether pointnet++ was every applied to semantic3d benchmark? Does anyone have any updates?