# 2st Multimodal Italian Sign Language Recognition C — Data

来源：https://www.kaggle.com/competitions/cvpr-mslr-2026-track-2/data

本地 Meta Kaggle 快照；不是本次实时页面。

# Dataset Description

## Directory Structure

```python
Sessioni CHALLENGE CVPR MSLR 2026/
├── train/ (117 sessions)
│ ├── 0_CLASSNAME/
│ │ ├── SAMPLE_{id}/
│ │ │ ├── SAMPLE{id}_RTM1.npy
│ │ │ ├── SAMPLE{id}_RTM2.npy
│ │ │ └── SAMPLE{id}_RTM3.npy
│ │ ├── ...
│ ├── 1_CLASSNAME/
│ │ ├── SAMPLE_{id}/
│ │ │ ├── SAMPLE{id}_RTM1.npy
│ │ │ ├── SAMPLE{id}_RTM2.npy
│ │ │ └── SAMPLE{id}_RTM3.npy
│ │ ├── ...
│ └── ... up to 125_CLASSNAME/
├── val/ (39 sessions; labels withheld)
│ ├── SAMPLE_{id}/
│ │ ├── SAMPLE{id}_RTM1.npy
│ │ ├── SAMPLE{id}_RTM2.npy
│ │ └── SAMPLE{id}_RTM3.npy
│ ├── ...
└── test/ (39 sessions; labels withheld)
├── SAMPLE_{id}/
│ ├── SAMPLE{id}_RTM1.npy
│ ├── SAMPLE{id}_RTM2.npy
│ └── SAMPLE{id}_RTM3.npy
├── ...
```

## What’s Inside Each Folder

- **Train set (`train/`)**
  - Organized by class:
    - `train/{classID_CLASSNAME}` → class ID `{classID}`, human-readable name
    - Inside each class folder: `SAMPLE_{id}/` subfolders

- **Validation (`val/`) and Test (`test/`) sets**
  - No class-level folders, each split contains only sample folders:
    - `val/SAMPLE_{id}/`
    - `test/SAMPLE_{id}/`
  - `id` is a globally unique identifier for each recording:
    - **12,600** IDs in train (100 sessions x 126 classes)
    - **5,040** in val (40 sessions x 126 classes)
    - **7,560** in test (60 sessions x 126 classes)
  - The assignment of `id`s to splits is fixed but randomized: **you cannot infer split membership from the ID alone**.

- **Files per sample folder (`SAMPLE_{id}/`)**
  1. **Radar Range-Time Maps**:
     - `SAMPLE_{id}_RTM1.npy` (antenna 1)
     - `SAMPLE_{id}_RTM2.npy` (antenna 2)
     - `SAMPLE_{id}_RTM3.npy` (antenna 3)

## What You Are Predicting

- **Task**: For each sample in **valtest/**, predict the correct gesture class ID (0-125).
- **Output**: A CSV with columns `id,Pred` that assigns each anonymized samples label to one of 126 classes. No partial results are allowed.

## Naming Conventions & Acronyms

- **CLASSNAME** - the human-readable name of the gesture (only in train)
- **RDM** - Range-Time Map from the 60 GHz radar (one file per antenna)

## Note

Feel free to explore the folders to verify file counts and formats. If anything looks unexpected, ask your questions in the discussion forum!