BdSL40 Dataset

This Bangla Sign Language Dataset BdSL40 comprises of 611 videos over 40 BdSL words with 8 to 22 video clips per word.

AI for Bangla 2.0

This project received an Honorable Mention at the AI for Bangla 2.0 Competition, under team name -- Silent Sentinels.

Dataset Link

BdSL40 Video
BdSL40 Frames

ISL to BDSL Mapping

ISL-to-BDSL. This mapping was created according to BdSL Dictionary and ISL Dataset. It is relatively easy to reproduce BdSL40 from ISL by following the mapping.

Kaggle Notebook

Training_on_BdSL40

Dataset Properties

Criterion	Count
Videos	611
Words	40

Label	Count		Label	Count
new	21		bad	21
lawyer	14		teacher	14
yesterday	14		time	15
friend	20		i	21
you	21		telephone	14
ring	14		winter	14
brown	21		skirt	19
pant	20		shoes	20
camera	14		heavy	8
soap	14		book	14
india	21		quiet	21
deaf	8		rich	8
thick	8		money	14
cow	21		fulfill	8
cheap	8		straight	8
life	8		more	21
crane	20		shirt	20
noon	14		bed	14
square	14		glad	8
tortoise	20		student	14

Samples

Label: Student

Label: Tortoise

Preprocessing

VideoResnet requires videos to be sampled into images and cropped to a square ratio. 32 frames are sampled from each video and the height and width are set to 100 pixels by default.

Usage:

python Preprocessing/preprocessing.py [path_to_dataset] [path_to_save_images]

Hyperparameters

The following hyperparameters are set by default.

"num_epochs": 120,
"learning_rate": 5e-5,
"batch_size": 64,
"h": 100,
"w": 100,
"mean": [0.5, 0.5, 0.5],
"std":  [0.5, 0.5, 0.5],
"total_frames": 32,
"start_skip": 6,
"end_skip": 8,
"batch_size": 64,
"num_classes": 40,
"test_ratio": 0.2

Training

Usage:

python Training/training.py [path_to_save_images]

Results after Training

Accuracy 82.93%

Acknowledgements

This dataset is adapted from by INCLUDE by Sridhar et al under the Creative Commons 4.0 license.