Article complete

TikTok Releases Depth Anything, Foundational Model for MDE

The dataset contains a total of 62 million images.

A dancer twirls in a flowing dress, with visual effects showcasing depth and layering. Background features a staircase.

Key takeaways

  • TikTok launches Depth Anything, a foundational model for Monocular Depth Estimation using 62 million images.
  • Achieve advanced depth estimation capabilities with zero-shot performance surpassing existing models like MiDaS and ZoeDepth.
  • Utilize a data engine for collecting and annotating vast datasets, enhancing model generalization.
  • Focus on simplicity and effectiveness rather than introducing new technical modules for image processing.
  • Implement challenging optimization targets and auxiliary supervision to strengthen model representations and semantic understanding.

Takeaways drafted with AI from the article text.

TikTok has unveiled Depth Anything, a groundbreaking development in the realm of Monocular Depth Estimation (MDE). This innovation harnesses the potential of a colossal dataset, comprising 1.5 million labeled images and an astonishing 62 million-plus unlabeled images. 

By jointly training on such a massive scale, Depth Anything emerges as a foundational model for MDE with an array of advanced features.

Key features of Depth Anything include 

  • Zero-shot relative depth estimation, surpassing MiDaS v3.1 (BEiTL-512)
  • Zero-shot metric depth estimation, outperforming ZoeDepth
  • Optimal in-domain fine-tuning and evaluation on NYUv2 and KITTI datasets

Unlike previous approaches, the focus of Depth Anything is not on introducing novel technical modules. Instead, the emphasis lies on constructing a straightforward yet potent foundational model capable of handling diverse images in any scenario.

To achieve this, the dataset is scaled up significantly through the implementation of a data engine designed for collecting and automatically annotating a vast pool of unlabeled data, totaling approximately 62 million images. This extensive dataset expansion proves instrumental in reducing generalisation errors.

Two effective strategies are explored in the process:

  • Challenging Optimisation Target: A more demanding optimization target is created using data augmentation tools. This compels the model to actively seek additional visual knowledge, thereby acquiring robust representations.
  • Auxiliary Supervision: An auxiliary supervision is developed to ensure the model inherits rich semantic priors from pre-trained encoders. This enhances the model's ability to interpret and understand images.

Extensive evaluation of Depth Anything's zero-shot capabilities involves six public datasets and randomly captured photos, showcasing its impressive generalisation ability. 

Furthermore, through fine-tuning with metric depth information from NYUv2 and KITTI, Depth Anything establishes new State-of-the-Art (SOTA) benchmarks. The improved depth model also yields superior results in depth-conditioned ControlNet.

Read: This New AI tool Could Mark the Beginning of the End for TikTok and Instagram Influencers

Published

Spotted an error? Tell our editors. Read our editorial standards and corrections.

More in this story

tiktok · 5 stories
  1. Voiceverse, the world’s first voice NFT, goes out of stock in 10 minutes
  2. You are here: TikTok Releases Depth Anything, Foundational Model for MDE
  3. TikTok's Parent Teases Video AI Model Rivaling OpenAI’s Sora, Turns Photos into Videos
  4. US and China Strike TikTok Deal, Oracle Steps in as Tech Partner
  5. Oracle and Silver Lake Take Control of TikTok’s US Business
About the author

Journalist

Mohit writes about AI in simple, explainable, and often funny words. He's especially passionate about chatting with those building AI for Bharat, with the occasional detour into AGI.

All stories
Belamy · Daily

The day’s AI news, before your first meeting.