recodai-luc-scientific-image-forgery-detection

[austere mirage]

Here is a quick analysis of the data and the problem:
https://www.kaggle.com/code/isaacmenard/scientific-image-forgery-data-analysis

[barren lily]

Hello everyone, and a warm welcome to the competition!

This is a milestone we've been working on for a long time.

Good luck, and thank you for joining this important mission!

[sinful cairn]

Hi, I noticed that the test dataset for the “ReCodAI LUC Scientific Image Forgery Detection” competition contains only one image. Could you please confirm if this is correct, or if there was an issue with the upload? It would be great if you could update the dataset or clarify the situation. Thanks!

[sterile trellis]

It contains many for training and just one toy example for test. Real test remains hidden.

[barren lily]

Hi everyone,
The test set is correct. It has more than one image and is private. It is only accessible during submission, meaning that during submission, your notebook will have access to the test_images directory containing all the private data.
Double-check the code requirements section for more info: https://www.kaggle.com/competitions/recodai-luc-scientific-image-forgery-detection/overview/code-requirements

There is an unofficial example from a participant that might help you with the submission. https://www.kaggle.com/code/isaacmenard/submission-scientific-image-forgery-detection
(thanks @austere mirage 👏 )

[thin cave]

How do i find helpful research materials for the competition?

[austere mirage]

https://www.kaggle.com/competitions/recodai-luc-scientific-image-forgery-detection/discussion/613066

here someone sent ressources and methods!

[eager verge]

Is anyone here testing with subtractive masking as a part of image preprocessing? Seems like a decent amount of FP from pretrains fine tuned on the dataset around non-organic imagery, especially when they also include gel stains.

Starting to experiment with mask generation around text and graphs first, subtracting those areas from the training set, and then training the last-mile model to generate a mask that identifies copy-move.

[green shoal]

Can it be like multiple different segments are duplicated rather than only one?
It’s possible right? But does that happen usually?

[sterile trellis]

At least at train they are, masks have N channels. Each channel for each object. About how often, there are EDA puplic notebooks.

[sinful raft]

Dear All,

Good morning and a warm welcome to everyone.

We are excited to announce the collaboration with Think Lab, marking a new step towards innovation, research, and technology-driven learning. This partnership aims to foster creativity, hands-on experimentation, and collaborative problem-solving among our students and faculty.

Together, we look forward to exploring new opportunities, sharing ideas, and building impactful projects that align with our vision of excellence and innovation.

Let’s make this collaboration a great success!

[gentle lagoon]

mr.pradeepvazhakad_45554 has been warned

Reason: Posted an invite

[viscid heath]

Hi

[barren lily]

Hi everyone,

In case you missed it, we have added supplemental training data featuring more complex structures collected from real-world forgeries. These new images were acquired and labeled using the exact same process that will be used for the final test set.

[south wedge]

hi guys, can i use dino v3 in my solution ??

i am bit confused as it has custom license (https://ai.meta.com/resources/models-and-libraries/dinov3-license/) . Could anyone help me out.

[barren lily]

Yes, you can use it for this competition.

[south wedge]

thank you!

[lusty magnet]

Hello, should the current submission.csv only contain one item ? for the single image that was given. ty.
I am also wondering if its the kaggle notebook that needs to generate the submission.csv

[lusty magnet]

as can be seen here, only one image https://www.kaggle.com/code/vinothkumarsekar89/dummy-submission-all-authentic

[sweet widget]

hello, can someone help me with a submission - I am receiving "Submission Scoring Error ". I have checked output, should be ok, I only run on the test image..

[sterile trellis]

Hi. Make sure your code is general for any number of images and ids that can be found in hidden test. That single test image is just an example.

@sweet widget so first read and store ids, then process them, and last write submission based on those hidden ids

[sweet widget]

@sterile trellis if we are talking from RAM (memory) perspective, probably the best approach is to write intermediary the results to csv (flush them as fast), not to keep RLE strings into memory

Also - @sterile trellis another topic, it seems weird, when I run notebook for me it takes less than 2 min...(on Kaggle with GPU T4 accelor) - but when running the submission...takes...a lot of time

[tardy talon]

thier are full hidden set of images that take some time to process and for submission error use ',' between integer and don't change the 'id' to int

[mystic hornet]

Nothing will be updated in the next four month?