Toilet Sign Matching

Restroom Pictogram Matching

Problem Description

This task focuses on training a model to learn the relationship between male and female restroom pictograms from the same restroom. Based on labeled training data, you need to train a matching model that, given a query image, finds its different-gender counterpart (for example, matching a male pictogram from the cropped image with the female counterpart from the original image), with the constraint that both pictograms come from the same restroom.

For example, for a query image (i.e., a male pictogram in the cropped image) shown below in Fig. 1, its corresponding counterpart is shown in Fig. 2 (i.e., a female pictogram in the original image, which is from the same restroom).

Figure 1	Figure 2

Dataset

Training Set

This dataset can be used for training the model. Directory structure:

1train/ 2├── crop/ # Cropped pictograms 3│ ├── female/ 1.png, 2.png, ... 4│ └── male/ 5└── orig/ # Original pictograms6    ├── female/7    └── male/

Each subfolder contains pictograms named from 1.png to 82.png, where the number indicates the restroom ID. For each restroom, there are four images (note that four images from the same restroom will have the same unique ID across all four subfolders):

• crop/female/i.png → Cropped female pictogram
• crop/male/i.png → Cropped male pictogram
• orig/female/i.png → Original female pictogram
• orig/male/i.png → Original male pictogram

Validation and Test Sets

The validation set (validation_set) and test set (test_set) contain two subfolders with the following structure:

1validation_set/2├── query/ # Cropped pictograms to be matched 3└── gallery/ # Candidate original pictograms

1test_set/ 2├── query/ # Cropped pictograms to be matched 3└── gallery/ # Candidate original pictograms

where

• query/ - cropped pictograms that need to be matched;
• gallery/ - set of original pictograms to match from;
• All images are in .png format.
• The validation set has 10 images in the query/ folder, and the test set has 30 images in its query/ folder.

Notes:

• Unlike the proposed split for the training set, the file names in query/ and gallery/ are numbered and shuffled independently, meaning matching cannot be based on IDs.
• For each cropped pictogram in query/ there are exactly two originals in gallery/ (one male, one female from the same restroom), meaning len(gallery)=2*len(query).

Task

For each image in the query/ folder, predict the image from gallery/ that:

• Is of opposite gender and comes from the same restroom.

This matching should be performed using the trained model.

Submission File

You need to generate a submission.csv file containing the following:

1image_path, label2dataset/validation_set/query/1.png, 73dataset/test_set/query/1.png, 24...

where:

• image_path is the path to the image from query/ for which the prediction is made (taken from test.csv)
• label is the number contained in the title of the image from gallery/ that matches with the one from query.

Score

The evaluation metric will be classification accuracy, defined as the proportion of correctly predicted samples out of the total number of evaluated samples.

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

• accuracy ≥ 0.9 → 100 points
• accuracy ≤ 0.5 → 0 points
• For intermediate values, proportional scoring between 0 and 100 is awarded.