High-Definition Segmentation in Google Meet – Google AI Blog

In latest years video conferencing has performed an more and more vital position in each work and private communication for a lot of customers. Over the previous two years, we’ve got enhanced this expertise in Google Meet by introducing privacy-preserving machine studying (ML) powered background options, also called “virtual green screen”, which permits customers to blur their backgrounds or substitute them with different photos. What is exclusive about this answer is that it runs immediately within the browser with out the necessity to set up extra software program.

So far, these ML-powered options have relied on CPU inference made doable by leveraging neural community sparsity, a typical answer that works throughout units, from entry stage computer systems to high-end workstations. This allows our options to achieve the widest viewers. However, mid-tier and high-end units typically have highly effective GPUs that stay untapped for ML inference, and current performance permits internet browsers to entry GPUs through shaders (WebGL).

With the newest replace to Google Meet, we are actually harnessing the ability of GPUs to considerably enhance the constancy and efficiency of those background results. As we element in “Efficient Heterogeneous Video Segmentation at the Edge”, these advances are powered by two main elements: 1) a novel real-time video segmentation mannequin and a couple of) a brand new, extremely environment friendly strategy for in-browser ML acceleration utilizing WebGL. We leverage this functionality to develop quick ML inference through fragment shaders. This mixture ends in substantial good points in accuracy and latency, resulting in crisper foreground boundaries.

CPU segmentation vs. HD segmentation in Meet.

Moving Towards Higher Quality Video Segmentation Models
To predict finer particulars, our new segmentation mannequin now operates on excessive definition (HD) enter photos, reasonably than lower-resolution photos, successfully doubling the decision over the earlier mannequin. To accommodate this, the mannequin have to be of upper capability to extract options with enough element. Roughly talking, doubling the enter decision quadruples the computation price throughout inference.

Inference of high-resolution fashions utilizing the CPU just isn’t possible for a lot of units. The CPU could have just a few high-performance cores that allow it to execute arbitrary complicated code effectively, however it’s restricted in its skill for the parallel computation required for HD segmentation. In distinction, GPUs have many, comparatively low-performance cores coupled with a large reminiscence interface, making them uniquely appropriate for high-resolution convolutional fashions. Therefore, for mid-tier and high-end units, we undertake a considerably quicker pure GPU pipeline, which is built-in utilizing WebGL.

This change impressed us to revisit among the prior design selections for the mannequin structure.

• Backbone: We in contrast a number of widely-used backbones for on-device networks and located EfficientNet-Lite to be a greater match for the GPU as a result of it removes the squeeze-and-excitation block, a element that’s inefficient on WebGL (extra beneath).
• Decoder: We switched to a multi-layer perceptron (MLP) decoder consisting of 1×1 convolutions as a substitute of utilizing easy bilinear upsampling or the costlier squeeze-and-excitation blocks. MLP has been efficiently adopted in different segmentation architectures, like DeepLab and PointRend, and is environment friendly to compute on each CPU and GPU.
• Model dimension: With our new WebGL inference and the GPU-friendly mannequin structure, we have been in a position to afford a bigger mannequin with out sacrificing the real-time body fee needed for easy video segmentation. We explored the width and the depth parameters utilizing a neural structure search.

HD segmentation mannequin structure.

In combination, these adjustments considerably enhance the imply Intersection over Union (IoU) metric by 3%, leading to much less uncertainty and crisper boundaries round hair and fingers.

We have additionally launched the accompanying mannequin card for this segmentation mannequin, which particulars our equity evaluations. Our evaluation reveals that the mannequin is constant in its efficiency throughout the assorted areas, skin-tones, and genders, with solely small deviations in IoU metrics.

Accelerating Web ML with WebGL
One frequent problem for web-based inference is that internet applied sciences can incur a efficiency penalty when in comparison with apps operating natively on-device. For GPUs, this penalty is substantial, solely attaining round 25% of native OpenGL efficiency. This is as a result of WebGL, the present GPU commonplace for Web-based inference, was primarily designed for picture rendering, not arbitrary ML workloads. In explicit, WebGL doesn’t embody compute shaders, which permit for normal goal computation and allow ML workloads in cellular and native apps.

To overcome this problem, we accelerated low-level neural community kernels with fragment shaders that sometimes compute the output properties of a pixel like colour and depth, after which utilized novel optimizations impressed by the graphics group. As ML workloads on GPUs are sometimes sure by reminiscence bandwidth reasonably than compute, we targeted on rendering strategies that may enhance the reminiscence entry, reminiscent of Multiple Render Targets (MRT).

MRT is a characteristic in fashionable GPUs that permits rendering photos to a number of output textures (OpenGL objects that characterize photos) without delay. While MRT was initially designed to help superior graphics rendering reminiscent of deferred shading, we discovered that we might leverage this characteristic to drastically scale back the reminiscence bandwidth utilization of our fragment shader implementations for essential operations, like convolutions and absolutely related layers. We accomplish that by treating intermediate tensors as a number of OpenGL textures.

In the determine beneath, we present an instance of intermediate tensors having 4 underlying GL textures every. With MRT, the variety of GPU threads, and thus successfully the variety of reminiscence requests for weights, is decreased by an element of 4 and saves reminiscence bandwidth utilization. Although this introduces appreciable complexities within the code, it helps us attain over 90% of native OpenGL efficiency, closing the hole with native purposes.

Left: A basic implementation of Conv2D with 1-to-1 correspondence of tensor and an OpenGL texture. Red, yellow, inexperienced, and blue containers denote completely different places in a single texture every for intermediate tensor A and B. Right: Our implementation of Conv2D with MRT the place intermediate tensors A and B are realized with a set of 4 GL textures every, depicted as pink, yellow, inexperienced, and blue containers. Note that this reduces the request rely for weights by 4x.

Conclusion
We have made speedy strides in enhancing the standard of real-time segmentation fashions by leveraging the GPU on mid-tier and high-end units to be used with Google Meet. We stay up for the chances that will likely be enabled by upcoming applied sciences like WebGPU, which carry compute shaders to the online. Beyond GPU inference, we’re additionally engaged on enhancing the segmentation high quality for decrease powered units with quantized inference through XNNPACK WebAssembly.

Acknowledgements
Special because of these on the Meet workforce and others who labored on this undertaking, particularly Sebastian Jansson, Sami Kalliomäki, Rikard Lundmark, Stephan Reiter, Fabian Bergmark, Ben Wagner, Stefan Holmer, Dan Gunnarsson, Stéphane Hulaud, and to all our workforce members who made this doable: Siargey Pisarchyk, Raman Sarokin, Artsiom Ablavatski, Jamie Lin, Tyler Mullen, Gregory Karpiak, Andrei Kulik, Karthik Raveendran, Trent Tolley, and Matthias Grundmann.