Pre-trained model shows low performance
KosukeArase opened this issue · 10 comments
Hi @laughtervv
Thank you for open sourcing your great work!
I downloaded your pre-trained model from here (the link in README) and your h5 data from here (the link in issue#3) and used pergroup_thres.txt and mingroupsize.txt in issue#8.
However, I ran test.py using them and got mAP 0.47, this is lower than the mAP reported in your paper (mAP 0.5435).
Did I make some mistakes?
Here is the result I got. Thank you.
Instance Segmentation AP: [ 0.47166667 0.53958333 0.37652931 0.67391304 0.43280553 0.7375
0.67703333 0.42900336 0.34550953 0.2 0.38823529 0.51387205
0.32641463]
Instance Segmentation mAP: 0.470158929369
Semantic Segmentation IoU: [ 0.85866874 0.92186948 0.86302137 0.84426813 0.72262609 0.8738558
0.89429596 0.82030826 0.79822212 0.72040915 0.7932732 0.74137266
0.81986088]
Semantic Segmentation Acc: %f 0.7979087289103282
@laughtervv Can you share the list of training data?
The result above is evaluated on randomly chosen scenes.
However, when I evaluated on Area_5, the mAP was 0.6447, which is much higher than reported in you paper.
Thank you.
Hi @KosukeArase ,I'm confused about how to write the "data/train_hdf5_file_list.txt".Is it like the file "filename.txt" given by @laughtervv ? I rename "filename.txt" to "train_hdf5_file_list.txt" to run train.py. But this caused my computer to be full of memory and eventually crashed. How can I get a start?
And I find "filename.txt" has a lot of duplicate names. What does it mean?
THANKS!
Hi @lhiceu
The program loads all data onto your memory, so it requires a few GB memory.
If you have enough memory, i think it is because the duplication.
I'm not sure what is filename.txt, but I simply created a list of h5 files which has no duplication.
Thanks for your reply! @KosukeArase
I also make a list of h5 file. Now I'm runing train.py with pretrained model given by the author, although the loss looks a little strange(Probably not changing much). I'll wait for a result.
By the way, did you pretrain the network as the paper said at Section 3.3 The network is trained with only the L_sim loss for the first 5 epochs?
I modify the get_loss() in model.py and epoch= 5,is_training = Truein train.py. But I find the trained model has something wrong to load. So I really don't understand how to do it.
I'm so sorry for too many problems.
THANKS!
Hi @lhiceu @KosukeArase
Now I am thinking about how generate the "train_hdf5_file_list.txt" that the program requires, with which I have no ideas yet. I dont understand what is a so-called "hdf5_file_list" file. I tried to find the file called filename.txt that you have mentioned above, however, I couldnt find it out.
Fortunately, you said that you had already made a list of h5 file by your own. I hope that you can help me generate a list of h5 file. If you are willing to share your list of h5 file by, for example, attaching a link to your comment for downloading the file, that would be so much help to me.
Hope that you are willing to help me. THANKS SO MUCH!!
@SinestroEdmonce
I download filename.txt from a issue reply that I forgot. SORRY.
But I think this file is wrong, because of its too many duplicate names. So I write a train_hdf5_file_list.txt for training. It is a list of .h5 file names without repetition.
@SinestroEdmonce
Here is the code to generate train_hdf5_file_list.txt
import os,shutil
source_path = './data/train/' #train file
f = open('train_hdf5_file_list.txt','w')
room_list = os.listdir(source_path)
for room in room_list:
f.write(room + '\n')
f.close()
@lhiceu
THANKS SO MUCH!!
It is so helpful for me!
Hi @lhiceu
Have you ever run the program? Which dataset did you use to do the testing, ModelNet40, ShapeNet or ScanNet? I really wanna run the program, which is of much importance for me. However I don't know how to make .h5 files from those datasets that I mentioned above. Besides, if you have already run the program, would you please tell me the procedures? It will be so nice of you to give me some guidance!
THANKS and appreciate all your help!
Hi @lhiceu
Have you ever run the program? Which dataset did you use to do the testing,ModelNet40,ShapeNetorScanNet? I really wanna run the program, which is of much importance for me. However I don't know how to make.h5files from those datasets that I mentioned above. Besides, if you have already run the program, would you please tell me the procedures? It will be so nice of you to give me some guidance!
THANKS and appreciate all your help!
hello,have you resolve the problem——“Which dataset did you use to do the testing, ModelNet40, ShapeNet or ScanNet?”.I'm confusing about this program and meet the same trouble .