🔢🧮 The archive of a thousand cameras 📚📷

📖 Description 🌌

At the heart of a secret research institute lies The archive of a thousand Cameras, an underground labyrinth with chambers monitored 24/7 by smart cameras. Each room captures objects, furniture, electronic devices, or symbols of mysterious experiments.

For unknown reasons, some of the archive data have lost their numerical labels: recordings show the question “how many objects of type X are in the image?”, but the answer is missing. Researchers tried to manually reconstruct the information, but thousands of images are impossible to process without help.

This is where you step in.

You have been entrusted with the role of a “Numerical Visual Analyst”, the only person authorized to build a multimodal model capable of looking at an image, understanding the question, and correctly answering with a single number.

Your mission is to use your expertise in Artificial Intelligence algorithms to restore the archive.

---

📁 Dataset Structure

There are two main files and a folder with images.

1. train.csv — Training Data

sample_id	question	answer	image_id
TRAIN0000	how many chairs are there	6	image888
TRAIN0001	how many pictures are on the wall	2	image942
TRAIN0002	how many chairs are there	10	image1404

2. test.csv — Data without answers

sample_id	question	image_id
TEST0000	how many drawers are there	image1362
TEST0001	how many televisions are there	image169
TEST0002	how many flags are visible	image1326

3. Images directory

images/
image888.png
image942.png
image1404.png
image1362.png
image169.png
image1326.png

📤 Submission Format

The submission.csv file should have the following format:

sample_id	answer
TEST0000	4
TEST0001	1
TEST0002	3

---

🎯 Evaluation Metric

Since the answers are numbers, the official evaluation uses MAE:

The final score is calculated based on the MAE obtained using the following rules:

• MAE ≤ 0.2 → 100 points
• MAE ≥ 1 → 0 points
• For intermediate values, the score is scaled proportionally between 0 and 100.