search

NadaAbbasMohamed / sonar-fish-imagery-dataset

5 stars 1 forks source link

About dataset #1

Open ShaofengZou opened 3 years ago

ShaofengZou commented 3 years ago

Hello, 

  Thanks for your sharing.  I see some speckle noise in the training set.  Can your provide the original images or video of the training set?  And can you share the original dataset url ? That's would be much help.

 Looking forward to your reply.
NadaAbbasMohamed commented 3 years ago

Hi this is the link for the original images with their annotations https://drive.google.com/file/d/1mbQ7KPNUeMz4t-KpYdTcBDkMoF3C3usP/view?usp=sharing I couldn't appload the whole dataset as a whole chunk due to the upload size limitations. this is the link to the original site where I got the videos: http://www.soundmetrics.com/Image-Gallery/Fisheries?page=4&q=

ShaofengZou commented 3 years ago

Cool! Thanks for your reply.

Besides, i want to consult you on a few questions.

  1. I find the original videos in the site you supply are in GIF format. Do you extract each frame of these GIFs and then label them?
  2. What is the difference between cluster_frame.jpg and fish_frame.jpg?
  3. The entire dataset seems to have 12,078 frames with annotation. What a massive work to label these images! Is that all done by you on hand?
  4. I find some labels in the text file are 1, but most label are 0. Is these label with 1 are wrong labels?
  5. Besides the dataset you share, what about the algorithm you will use to detect fish? Do you have a plan to write a paper about this to share your contribution?
NadaAbbasMohamed commented 3 years ago

You are welcome.

  1. Yes I extracted the images from the GIFs, cleaned them up from repeated frames, done some enhancements on the images as sharpening them alittle and then I labelled them. The speckle noise was added to augment the images.
  2. In some images you will find clusters of fish. It was nonsense for me to label each fish in that cluster so I assigned it to be another separate class - this leads to question 4: 1 is the label for the cluster class and 0 is for the fish class. you can simply remove any annotation for the cluster if you want to. A simple code will do it for you.
  3. Yes
  4. Unfortunately I have no intention to continue on with the project at this time. But I can tell you that models as MobilNet and VGG didn't give me any good results (note I only trained the models on the fish images and didn't use those containing clusters)
ShaofengZou commented 3 years ago

Thank you for your detailed answer.:beer:

Maybe the fish is too small to detect. Enlarge the image then input maybe helpful. I will try later.

Can you upload the whole dataset to google driver or onedrive or baiduyu? :rose: That's would be much help for me to go deeper research. For google driver, you can zip the dataset into several parts to avoid the upload size limitations: .