samreenanjum / ctmc Goto Github PK

Hi and thanks for making this dataset!

While using the video frames, the JPEG blocking artifact (due to 8x8 px compression patches) appears to be much stronger than the original video:

import matplotlib.pyplot as plt
from torchvision.io import read_image, read_video

frame = 20

# video frames in the CTMC dataset
jpg = read_image(f"CTMCV1/train/PL1Ut-run03/img1/0000{frame}.jpg")
# uses ffmpeg to read the original video from Nikon website
mp4, _, _ = read_video("PL1Ut-run03.MP4")

f, ax = plt.subplots(1, 2, figsize=(10, 5))
ax[0].imshow(jpg[0, :80, :80], cmap="gray")
ax[1].imshow(mp4[frame, :80, :80, 0], cmap="gray")

This will likely become a confounding factor for training vision models, since entire patches are smoothed out in the JPEG files. However I do not see such severe artifacts in the paper. If both training and testing are both done based on the original MPEG-4 stream, is the test result still valid?

Hi @samreenanjum thanks for putting this dataset together! We are trying to reuse the dataset and would like to know imaging metadata such as the objective lens and the pixel size. If you do not have those at hand, can you point out the original source on Nikon's website?