metadata
license: mit
This is a indoor building element dataset with 37 classes.
This dataset consists of 900 original images collected from the ETH building videos collected by Aria glasses, Excel sheet: 230116_Ernest-Bloch_Inventaire and web-scraped samples from gruner-reuse.ch using the script parse_grunerweb.ipynb. The dataset is randomly split into 720 training, 90 validation, and 90 test images.
After applying data augmentations, the final dataset contains a total of 2,340 images, including:
- 2,160 augmented training samples
- 90 validation samples
- 90 test samples
Preprocessing
- Auto-Orient: Applied
- Resize: Stretched to 640×640
- Auto-Adjust Contrast: Using Adaptive Equalization
Augmentations
Each training example is augmented to generate 3 output variations with the following transformations:
- Hue: Between −15° and +15°
- Saturation: Between −25% and +25%
- Brightness: Between −15% and +15%
- Blur: Up to 1px
- Noise: Up to 0.1% of pixels
Coco format:
``` data_new_coco/ ├── train/ │ ├── images/ │ └── annotations/ ├── valid/ │ ├── images/ │ └── annotations/ ├── test/ │ ├── images/ │ └── annotations/ ├── README.dataset.txt └── README.roboflow.txt ```
Yolo format:
``` data_new_yolo/ ├── train/ │ ├── images/ │ └── labels/ ├── valid/ │ ├── images/ │ └── labels/ ├── test/ │ ├── images/ │ └── labels/ ├── data.yaml ├── README.dataset.txt └── README.roboflow.txt ```