ByteDance

Depth Anything V2 is a monocular depth estimation model released in June 2024 by researchers at the University of Hong Kong and TikTok. It predicts a dense depth map from a single RGB image, enabling 3D-aware applications without the need for stereo cameras, LiDAR, or multi-view inputs. The model improves on the original Depth Anything through three modifications: replacing real labeled images with 595K high-quality synthetic images during teacher training, scaling up teacher model capacity, and using the stronger teacher to generate pseudo-labels on 62 million unlabeled real images used to train the student models. This pipeline reduces the depth prediction artifacts that can occur in reflective, transparent, and texture-poor regions. Compared to diffusion-based depth models such as Marigold, Depth Anything V2 runs more than 10× faster while producing more accurate predictions.Depth Anything V2 is released in four sizes: Small (25M), Base (97M), Large (335M), and Giant (1.3B), and in two output modes: relative depth (normalized scene-level estimates) and metric depth (absolute distance in meters, produced by fine-tuning the relative-depth backbone on depth-annotated datasets). The Small, Base, and Large model weights are released under Apache 2.0, and the Giant variant under CC-BY-NC-4.0 for non-commercial use. A successor model, Depth Anything 3, was released in November 2025 by the ByteDance Seed team, extending the framework to multi-view depth estimation and camera pose recovery.

ByteTrack is a multi-object tracking algorithm developed by ByteDance, released in October 2021. It addresses a core limitation of earlier tracking methods by associating every detection box — not just high-confidence ones — with tracklets using a two-stage matching process. Low-score detection boxes, which typically correspond to occluded or blurry objects, are retained in a secondary matching step against unmatched tracklets, significantly reducing the number of lost tracks.ByteTrack achieves strong performance on the MOT17 and MOT20 benchmarks at the time of release, operating in real time on standard GPU hardware. It is architecture-agnostic and can be paired with any object detector, making it straightforward to integrate into existing detection pipelines for video analytics, surveillance, and sports tracking applications.