When working with video content programmatically, one common challenge developers encounter is handling embedded borders, watermarks, or overlays. These unwanted elements often degrade the quality and reusability of content, especially when videos are repurposed from one platform to another. This issue is particularly prevalent in content such as reposted TikToks, Instagram reels, or videos formatted for specific platforms — which have taken over the internet.
Sieve's new Border Detection function addresses this problem by automatically identifying and removing embedded borders from videos. This allows developers to streamline their workflows, improve video quality, and ensure content can be re-used without manual edits.
Below are some example cases of videos with unwanted borders automatically removed.
How do I use it?
To try a few sample videos, go to Sieve’s playground and upload a video. If you want to integrate this into your own product or workflow, you can also integrate the solution via API.
Evaluation and Limitations
We evaluated our solution against 1000 different reposted videos on the internet which included TikToks, post-edited embedded content, stream clips, and more. We measured accuracy to be how close our detected borders are compared to the ground truth. For instance, if the ground truth border is 100 pixels and our detection is 90 pixels, this results in an accuracy of 90%. Detection accuracy was greater than 95%, with failure cases and outliers tending to come from gradient borders, moving borders, or borders with visual color similarity to main content.
What’s next?
Reusability of internet content is a domain area we’re really interested in. We will continue to improve this solution while also building adjacent solutions that help developers understand and filter out more unwanted styles of videos as they automate video creation, curate large datasets for model training, and much more. In the meantime, we encourage you to experiment with what we are releasing today and reach out to contact@sievedata.com with any questions or feedback!