Researchers Release DeepLab2, a High-Performance TensorFlow Library for Deep Labelling Tasks

A team from Google Research, working alongside colleagues at the Technical University Munich, Johns Hopkins University, and KAIST, has released a new TensorFlow library targeting deep labelling: DeepLab2.

"DeepLab2 is a TensorFlow library for deep labelling," the team explains, "aiming to provide a unified and state-of-the-art TensorFlow codebase for dense pixel labelling tasks, including, but not limited to semantic segmentation, instance segmentation, panoptic segmentation, depth estimation, or even video panoptic segmentation.

"Deep labelling refers to solving computer vision problems by assigning a predicted value for each pixel in an image with a deep neural network. As long as the problem of interest could be formulated in this way, DeepLab2 should serve the purpose."

To prove DeepLab2's performance, the team used it with the Axial-SWideRNet as a network backbone and achieved a claimed 83.5 per cent mean intersection over union (mIoU) on the Cityscapes validation set, with only single-scale inference and ImageNet-1K pre-trained checkpoints.

DeepLab2 is already in use on a range of projects, with Panoptic-DeepLab, Axial-DeepLab, MaX-DeepLab, STEP or Motion-DeepLab, and ViP-DeepLab named in the documentation. The released code, however, isn't a one-to-one match to the paper describing it: "Note that this library contains our re-implemented DeepLab models in TensorFlow2," the team notes, "and thus may have some minor differences from the published papers."

"We hope that publicly sharing our library could facilitate future research on dense pixel labelling tasks and envision new applications of this technology."

The source code for DeepLab2 is available on GitHub under the permissive Apache 2.0 license; a preprint of the paper describing it is available on arXiv.org under open access terms.