Delivery

Model Accuracy

The following series of images are examples of images used to test the accuracy of the model we developed. Each image has a bounding box drawn for where the model identified an object, and what object it identified it as. Our general threshold for an acceptable identification is 70% confidence by the model, however this number is easily adjustable depending on customer needs. 

The model we developed was trained on 37 breeds of cats and dogs, as well as coyotes, foxes, raccoons, and squirrels. 

Model Speed

Through testing we found that each invocation of the model on an image took approximately 43 milliseconds. Our goal is to be able to perform detection in real time at either 25 or 30 frames per second, which would give a window of approximately  40 to 33.3 milliseconds to perform all necessary operations before the next frame arrives. 

The model uses 32 bit floating point tensors, so it is unsurprising that it doesn't quite meet the real-time requirements we were hoping for. This is where the use of the Coral EdgeTPU would be utilized, and is discussed further on the optimization page.  

Test Machine Specifications

Processor: AMD Ryzen™ 7 7840U

Memory: 64.0 GiB DDR5

Operating System: NixOS 24.05 (Uakari)

Linux Kernel Version: Linux 6.8.7