With the Coral Edge TPU™, you can run a semantic segmentation model directly on your device, using real-time video, at over 100 frames per second. You can even run a second model concurrently on one Edge TPU, while maintaining a high frame rate.
This page provides several trained models that are compiled for the Edge TPU, and some example code to run them.
These models are trained and compiled for the Edge TPU.
| Model name | Detections/Dataset | Input size | Depth mul. | Output stride | TF ver. | Latency 1 | Micro 2 | Model size | Downloads |
|---|---|---|---|---|---|---|---|---|---|
|
U-Net MobileNet v2 |
37 pets |
128x128x3 | N/A | N/A | 1 | 2.7 ms | Yes | 7.2 MB | |
|
U-Net MobileNet v2 |
37 pets |
256x256x3 | N/A | N/A | 1 | 29.0 ms | Yes | 7.3 MB | |
|
MobileNet v2 DeepLab v3 |
20 objects |
513x513x3 | 0.5 | N/A | 1 | 36.8 ms | No | 1.1 MB | |
|
MobileNet v2 DeepLab v3 |
20 objects |
513x513x3 | 1.0 | N/A | 1 | 43.0 ms | No | 2.9 MB | |
|
EdgeTPU-DeepLab-slim |
28 objects |
513x513x3 | 0.75 | N/A | 1 | 65.9 ms | No | 3.1 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
324x324x3 | 0.75 | 16 | 1 | N/A | Yes | 1.6 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
352x480x3 | 0.75 | 16 | 1 | 6.9 ms | Yes* | 1.6 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
512x512x3 | 0.75 | 16 | 1 | 10.7 ms | Yes* | 1.7 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
480x640x3 | 0.75 | 16 | 1 | 12.3 ms | Yes* | 1.8 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
576x768x3 | 0.75 | 16 | 1 | 17.7 ms | Yes* | 1.8 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
768x1024x3 | 0.75 | 16 | 1 | 30.8 ms | Yes* | 2.0 MB | |
|
MobileNet v1 BodyPix |
24 body parts |
720x1280x3 | 0.75 | 16 | 1 | 38.8 ms | Yes* | 2.3 MB | |
|
ResNet-50 BodyPix |
24 body parts |
288x416x3 | N/A | 16 | 1 | 46.9 ms | No | 24.5 MB | |
|
ResNet-50 BodyPix |
24 body parts |
480x640x3 | N/A | 16 | 1 | 384.0 ms | No | 26.6 MB | |
|
ResNet-50 BodyPix |
24 body parts |
496x768x3 | N/A | 32 | 1 | 87.0 ms | No | 26.9 MB | |
|
ResNet-50 BodyPix |
24 body parts |
624x864x3 | N/A | 32 | 1 | 153.5 ms | No | 28.5 MB | |
|
ResNet-50 BodyPix |
24 body parts |
672x928x3 | N/A | 16 | 1 | 737.2 ms | No | 35.3 MB | |
|
ResNet-50 BodyPix |
24 body parts |
736x960x3 | N/A | 32 | 1 | N/A | No | 38.6 MB |
1 Latency is the time to perform one inference, as measured with a Coral USB Accelerator on a desktop CPU. Latency varies between systems and is primarily intended for comparison between models. For more comparisons, see the Performance Benchmarks.
2 Indicates compatibility with the Dev Board Micro. Some models are not compatible because they require a CPU-bound op that is not supported by TensorFlow Lite for Microcontrollers or they require more memory than available on the board. (All models are compatible with all other Coral boards.)
* Although Dev Board Micro supports all the MobileNet v1 BodyPix models, beware that the on-board camera is 324x324 px, so you should use only the 324x324x3 model, unless you connect a larger-resolution camera.
If you want to process portrait-orientation images, download BodyPix models for portrait input.
Person segmentation with video
This example takes in a camera feed and performs body-part segmentation using the BodyPix model (with both MobileNet v1 and ResNet50 backbones). In addition to identifying different body parts, it can anonymize people from images.
Languages: Python
Semantic segmentation
This example performs semantic segmentation on an image. It takes an image as input and creates a new version of that image showing which pixels correspond to each recognized object.
Languages: Python