search

lyu-yx / ACUMEN

Official repo for “Unlocking Attributes' Contribution to Successful Camouflage: A Combined Textual and Visual Analysis Strategy”
8 stars 0 forks source link

How to generate my own Desc file #1

Open LaibinChang opened 1 month ago

LaibinChang commented 1 month ago

This is a very interesting work. How can I generate my own Desc_final file? In addition, this file cannot be downloaded “COS_data_all in_one.zip”. Looking forward to your reply.

lyu-yx commented 4 weeks ago

Hi there @LaibinChang ,

Apologies for the late reply, and thank you for your interest in our work. The updated COS_data_all_in_one.zip file can be found here.

Regarding the Desc_final file you mentioned, we used the GPT-4V API for initialization, followed by volunteer-based fine-tuning. If you'd like to create your own Desc_final file, you'll need to explore the appropriate prompts that suit your specific task. For your reference, I’ve included the prompt we used to generate COD-TAX.

Here is an image contains camouflaged object(s), answer the following questions one by one.

  1. Provide a detailed 30-word description of the image focusing on the camouflaged objects and their immediate environment.

  2. Describe how the camouflaged object(s) achieve camouflage. Please provide a detailed description within 30 words.

  3. Here are some camouflaged strategies or reasons which make this camouflage succuss. Surrounding Reasons: Background Matching, Surrounding Pattern Disruption, Environmental Motion Dazzle, Environmental Shading, Environmental Textures. Camouflaged Object-Self Reasons: Color Matching, Shape Mimicry, Behavior Mimicry, Texture, Shadow Minimization, Edge Diffusion. Imaging Quality Reasons: Blur Issue, Low Resolution, Improper Exposure, Compression Artifacts, Object Size Matters, Object Placement. Among them, Surrounding Reasons, Camouflaged Object-Self Reasons, and Imaging Quality Reasons are main classes while others are finer classes.

Calculate the exact contribution proportions of each finer class to the success of a camouflage in that image which make human hard to detect. Ensure that the total sum of these proportions equals 1. Include all relevant finer classes, even if their contribution is 0. Allocate proper contribution to the classes in “Imaging Quality Reasons” to make sure it is not always 0. Provide the results as a list, without any additional explanations. The focus is on precision and completeness in representing each finer class's contribution.

I hope this information helps!