tfmatch icon indicating copy to clipboard operation
tfmatch copied to clipboard
verified publisher icon lzx551402

The training set of ASLFeat

Open Jemmagu opened this issue 5 years ago • 8 comments

Hi @lzx551402 , thanks for your great work. You mentioned that your ASLFeat v2.0 is trained using GL3D. I'm curioused about the composition of this training set.

  1. For example, have you added the tourism dataset into the training set?
  2. Is this ASLFeat robust to illumination/weather/season variations? Have you evaluated that?
  3. I also wanna ask about what kind of improvement can be done on ASLFeat, like time cost or perfomrance improvement.

Best, quite looking forward to your reply.

Jemmagu avatar Apr 26 '20 08:04 Jemmagu

  1. yes
  2. yes, we have tested asl in the acchen dataset, which has large illumination change
  3. we believe the performance of asl can be further improved, but currently we also do not have answers ...

zjhthu avatar May 13 '20 02:05 zjhthu

Thanks, @zjhthu, Have some further questions:

  1. about the training set. why don't you use Megadepth dataset as used in D2net?
  2. about the backbone. Have you tried using VGG16 rather than L2net as backbone? Why you choose L2net?
  3. Just wanna ask have you tested how does aslfeat perform in indoor scenes?

Hope to get your reply, big thanks :)

Jemmagu avatar May 25 '20 08:05 Jemmagu

  1. GL3D is a much larger dataset than Megadepth, which is suitable for training a general-purpose local feature.
  2. L2Net is much smaller than VGG16, which is suitable for real-time applications.
  3. ASL is tested on FM-Bench in the paper, which contains one indoor dataset, the TUM dataset.

zjhthu avatar May 25 '20 08:05 zjhthu

Hi @zjhthu is the geolabel/corr.bin and img_kpts/<img_idx>.bin a must to train aslfeat?

Jemmagu avatar May 28 '20 02:05 Jemmagu

No, for aslfeat, they are not must.

zjhthu avatar May 28 '20 12:05 zjhthu

Hi @zjhthu , sorry to bother you agian I'm confused with the depht_image size. I loaded the pfm and found the depth array is 250250, then it seems be upsampled to 480480, which is same with the resized db image size.

But why the original depth size is 250? Because to my undetstanding ,it should be the same with ori_db_size (namely 1k*1k). And is it reasonable to upsample the depth?

Jemmagu avatar May 30 '20 03:05 Jemmagu

To reduce the storage of depth, the depth map size is set to 250*250. What do you mean by "reasonable"? Do you mean upsampling will cause errors at the depth border?

zjhthu avatar May 31 '20 14:05 zjhthu

Hi @zjhthu, I'm confused with your GL3D image size. In the cameras.txt, I saw the original image size is about 4k3k, but the images in the undist_images folders are all 1k1k. So did you directly crop the original image to 1k*1k, or you firstly crop it then resize it? wanna know some details about preprocessing the aslfeat training set.

Thanks a lot :)

Jemmagu avatar Jun 05 '20 08:06 Jemmagu