rmitra / ps-dataset Goto Github PK

May I ask where the camera intrinsics are stored? I couldn’t find them. Thanks so much.

Hi,

Could you please add to comparison on Fountain and Herzjesu HardNet++?
https://github.com/DagnyT/hardnet/tree/master/pretrained/pretrained_all_datasets

We cannot compare to HardNet-PS on HPatches, because HardNet++ was trained on Brown+HPatches, but can do on Fountain and Herzjesu. I think, it would be interesting to compare :)

Hi, can you tell me how to open the downloaded patchImage.bin file? Thank you.

Hi,

Thank you for putting link to the dataset! May I ask also for some downloading script to get full dataset?

Hello Rahul,
Thank you for contributing this dataset.

Question for you -- I am looking at scene 31, where I see the following lines

397, 67, 425.00,604.00,1.440,1.968
397, 166, 422.00,614.00,1.436,1.981
397, 135, 430.00,622.00,1.482,2.095
397, 152, 451.00,689.00,1.396,1.998

in patchinfo.txt.
When I render these images, and plot the corresponding points, they do not seem to be associated with the same 3d point. I am ignoring occlusion of course, which would mean they could still be the same 3d point, but visually we can't easily know this. Am I reading in the data incorrectly? If so, how can I take into account the occlusion when associating pixels?

Another example in Scene 31 is

336, 67,306,480,1.563,2.026,
336,135,302,494,1.574,2.227,
336, 40,317,516,1.574,2.200,
336,152,315,550,1.573,2.783,
336, 46,305,599,1.550,1.999,
336,117,314,566,1.547,2.919,
336,207,315,639,1.578,2.713,

It doesn't seem possible that these 2d points could be associated the same 3d point:

Thanks very much,
John

Hey, thanks for publishing your dataset.

In the pair_patch files, the pairs are written in this format: "3.2998e+05,3.2998e+05".
If I'm not wrong, this causes two problems:

• Some of the indices are out-of-range, as in scene 20 for example (last pair).
• The jumps are 10 patches, so we lose all the intermediate data (e.g 329981, 329982 ... ).

What do I miss?

Thanks!

Hello, I am interested in using your data set and am reading through the Training Procedure section (section 4.1) of your paper. I hope that you can clarify some details for me.

Selecting nearest non-matching pairs

In it you describe a modified process for selecting valid candidates for the nearest non-matching pair. It sounds like, to do this, I would need to:

• Find two 3D points from point_info.txt that have at least one image in common
• Compute the image-footprint of a patch for each of those 3D points, based on the patch's reported scaling and rotation
• See if the image footprints of the two patches from the respective 3D points overlap sufficiently (50% of the unnormalized patch size)
• If they do, record them as a "valid mismatch"

Does this seem like I have a correct description of the process? Similar to pair_patch.txt do you have a valid_mismatch.txt where you have already curated these patches?

To clarify your instructions further, when you say:

and their projections in that common image differ by 50%

Do you mean by at least 50% or at most 50%?

Could you indicated a valid pair of non-matching patches from, say, sequence 4 for me? This might help me have more confidence that I understand this properly.

Forming batches

Regarding the way that you compose batches for training:

• How do you ensure that there are a sufficient number of valid patches for nearest non-matching pairs in a given batch?
• Do you only draw from a single sequence to create a given batch?
• Do you have a sense for how important it is to use the valid-mismatch procedure in your training, as compared to the standard procedure in HardNet?

Thank you!

Thanks for publishing this paper! You guys did a great job describing your dataset creation algorithm and your PS-HardNet results are truly impressive 😄

Hey, I found PS-Dataset to be very helpful, do you have any plan to increase the size of the dataset? For example including more scenes?

This script looks don not download the image in the datasat, how to solve this problem? Thanks :)

Hi @rmitra I hope all is well.

I discovered that my model was doing anomalously-poorly on scenes 41, 49, and 71 from your data set.

After some probing, I discovered that these scenes contain a substantial number patch-pairs with relative rotations up to 180-deg - much larger than is documented, if I am not mistaken.

The following patches have not been augmented in any way.

Example from scene 41:

Example from scene 49:

Example from scene 71:

Hi, I've download the dataset and transform all the bin into the regular jpg file so that we can visualize the data. Here I have a question.

In the folder index 11, the 0th to the 3rd patch look like they belong to the same pair (as we can also confirm this by seeing the patch_info.txt file, which the first to fourth line's first index are all 1). But in the pair_patch.txt file, I only see (0,2), (0,3) which means the 0th patch is paried with the 2nd and 3rg patch, but patch 1 should also be paired.

Is there anything wrong in the pair_patch.txt?

Thanks.

Hi,
Is there any instruction on how to read the bin files?
Thanks!

I think your sh script does not work for me, since bashing the script says

Unable to establish SSL connection

Can you please check whether it is being supported or not ?

Hi, thanks for the dataset!

I am trying to parse the bin file and display the images inside, but I notice that there the forntal 321 or 323 bytes seems not image data, am I right? and what do those bytes mean?

do those image patches for folder 11 look correct ?

Below are my script to parse bin file

from PIL import Image
import numpy as np
import os
import pdb

folder = "11"
infoFileName = "patch_info.txt"
binFileName = "patchImg.bin"
infoFilePath = os.path.join(folder, infoFileName)
filePath = os.path.join(folder, binFileName)

numOfImgs = 0
with open(infoFilePath, "r") as datasetInfo:
    for line in datasetInfo:
        numOfImgs += 1
    print("folder", folder, "has", numOfImgs, "images")

width = 48
height = 48
with open(filePath, "rb") as dataset:
    fileSizeInByte = os.stat(filePath).st_size
    biasInByte = fileSizeInByte - numOfImgs*height*width
    dataset.read(biasInByte)
    print("binFile", binFileName, "has", biasInByte, "unknown bytes in header")

    image = np.zeros((height,width,3), dtype=np.uint8)
    try:
        # for i in range(biasInByte):
        #     print(ord(dataset.read(1)))

        for idx in range(5):
            for h in range(height):
                for w in range(width):
                    image[h,w,0] = ord(dataset.read(1))
                    image[h,w,1] = image[h,w,0]
                    image[h,w,2] = image[h,w,0]

            img = Image.fromarray(image, 'RGB')
            img.save(str(idx) + ".jpg")
    except:
        pass

Hey,

Can you please upload your pytorch code for processing the dataset and use it to train the network with HardNet loss?

Thanks!