In this episode of the Talking Papers Podcast, I hosted Chamin Hewa Koneputugodage to chat about OUR paper “DiGS: Divergence guided shape implicit neural representation for unoriented point clouds”, published in CVPR 2022.
In this paper, we took on the task of surface reconstruction using a novel divergence-guided approach. Unlike previous methods, we do not use normal vectors for supervision. To compensate for that, we add a divergence minimization loss as a regularize to get a coarse shape and then anneal it as training progresses to get finer detail. Additionally, we propose two new geometric initialization for SIREN-based networks that enable learning shape spaces.
I first met Chamin when he was still doing his honours. He would come to Steve’s group meetings and report his progress on vision and language. After he started his PhD and I saw he was looking for his next project I opened the door to 3D. The true story behind this paper is that it started off as my baby but because I had to move back to Israel, I recruited Chamin to help out with a few experiments (Thanks Steve!!!). After a short time, where Chamin worked out most of the maths for the initialization, ran all of the experiments for shape space, and had as much understanding and contribution as I had in the project, it became his baby as much as mine. I am proud to share this brain child with him.
Shape implicit neural representations (INR) have recently shown to be effective in shape analysis and reconstruction tasks. Existing INRs require point coordinates to learn the implicit level sets of the shape. When a normal vector is available for each point, a higher fidelity representation can be learned, however normal vectors are often not provided as raw data. Furthermore, the method’s initialization has been shown to play a crucial role for surface reconstruction. In this paper, we propose a divergence guided shape representation learning approach that does not require normal vectors as input. We show that incorporating a soft constraint on the divergence of the distance function favours smooth solutions that reliably orients gradients to match the unknown normal at each point, in some cases even better than approaches that use ground truth normal vectors directly. Additionally, we introduce a novel geometric initialization method for sinusoidal INRs that further improves convergence to the desired solution. We evaluate the effectiveness of our approach on the task of surface reconstruction and shape space learning and show SOTA performance compared to other unoriented methods.
LINKS AND RESOURCES
To stay up to date with Chamin’s latest research, follow him on:
Recorded on April 1st 2022.
If you would like to be a guest, sponsor or just share your thoughts, feel free to reach out via email: firstname.lastname@example.org
SUBSCRIBE AND FOLLOW
🎧Subscribe on your favourite podcast app: https://talking.papers.podcast.itzikbs.com
📧Subscribe to our mailing list: http://eepurl.com/hRznqb
🐦Follow us on Twitter: https://twitter.com/talking_papers
🎥YouTube Channel: https://bit.ly/3eQOgwP