github dbro/frigate-detector-edgetpu-yolo9 v1.5

latest releases: v1.7, v1.6
one month ago

Faster
Reduced detection time to 10ms (from 12ms). This is mostly the result of having the TPU calculate the maximum class score for each detection candidate, which saves the CPU some time in post-processing. There are also some efficiency gains from filtering as early as possible in the post-processing sequence of operations.

Larger
In addition to the model with input image size 320x320 already available, this version introduces a model with input image size: 512x512 pixels. The larger size model can detect smaller objects, for example when they are farther away. Detection time for the larger model is 21ms on my 12+ year old system, and I expect it would run a bit faster on more recent hardware.

UPDATE: see below for a 3rd model option, 320x320 "hardswish"

Compare the difference in detections (using a minimum score of 0.25, and NMS to remove duplicates) between the 320 and 512 image sizes:

The 320x320 version has 9 detections: 5 persons and 4 cars.
abbeyroad-detections-320

The 512x512 version has 12 detections: 5 persons, 6 cars, and 1 truck (which was detected as a car by the 320x320 model). The additional cars are smaller, in the middle of the image. The confidence scores for the people and cars are all higher than in the 320 version.
abbeyroad-detections-512

OBJECT CLASS Score according to 320x320 version Score according to 512x512 version
George person 0.84 0.88
Paul person 0.80 0.88
Ringo person 0.86 0.87
John person 0.73 0.84
bystander person 0.28 0.39
VW Beetle car 0.74 0.81
truck car/truck 0.51 0.62
car car 0.38 0.70
car car 0.37 0.61
car car <0.25 0.46
car car <0.25 0.31
car car <0.25 0.28

UPDATE
There is now another 320x320 model available for download. Its name refers to the "hardswish" activation function used. It achieved higher accuracy scores than the "relu6" version, 75 mAP vs 72 mAP. In practice, the detections are more confident, which might indicate more true positives; but maybe also increases the potential for false positives. It takes approximately twice the time to perform detection 18ms vs 10ms for relu6.

Don't miss a new frigate-detector-edgetpu-yolo9 release

NewReleases is sending notifications on new releases.