Skip to product information
1 of 13


Low-Light Dataset

A synthetic video dataset for fitness featuring a variety of challenging light scenarios

Regular price $0.00
Regular price Sale price $0.00
Sale Sold out

The Fitness Low-light Dataset focuses on various challenging lighting scenarios within home fitness and physical therapy (PT) environments. The dataset contains 3 different exercises being performed by diverse avatars in settings with low light and correspondingly low signal to noise ratio (SNR). The dataset introduces several output metrics that report brightness, contrast, and SNR associated with each frame. 



<li>120 videos across 3 exercises</li>
<li>Varying low-light scene illumination</li>
<li>Outputs chosen to challenge state-of-the-art keypoint detection models</li>
<li>Within the low-light illumination context, variety of brightnesses and color contrasts</li>
<li>Illumination-dependent noise simulation</li>
<li>No out-of-frame or occluded avatars</li>
<li>7-15 reps per video</li>
<li>30 FPS, 512x512 resolution videos</li>
<li>Variation in kinematic trajectories and exercise speeds</li>
<li>Avatars with diverse heights and body shapes</li>
<li>4 interior environments</li>


The dataset includes the following exercises:
<li>Bicep curls (alternating)</li>
<li>Sumo squats </li>


Each video is accompanied by a rich set of pixel-perfect labels and annotations. Highlights for this dataset:
<li>Mean brightness of entire 2D image</li>
<li>Mean brightness of avatar pixels</li>
<li>Mean brightness of local region surrounding the avatar</li>
<li>Brightness contrast between avatar and local surroundings</li>
<li>Color contrast between avatar and local surroundings</li>
<li>Color contrast between avatar and local surroundings.</li>
<li>Mean signal to noise ratio of entire 2D image</li>
<li>All the usual labels including rep counts, avatar characteristics, camera location, 2D/3D kyepoints, etc. </li>
<li>And much more!</li>
For the full description of labels and metadata, check out the <a href="" target="_blank">README</a>.


512 x 512 mp4 videos (30 FPS).
Total size: 6.0 GB
Each exercise is in a different zipped folder (1.0-2.6 GB per folder).


This dataset is licensed under Infinity AI’s <a href="" target="_blank">Terms and Conditions</a>.


<li>Github <a href="" target="_blank">README</a>: full dataset and annotation descriptions.</li>

Questions? We’re happy to chat asynchronously via email or hop on a call. Just send us a note at (this goes to all of the Infinity AI founders).

View full details