Dynamic, contextualized AI models
A method for employing a semi-supervised learning approach to improve accuracy of a small model on an edge device is presented. The method includes collecting a plurality of frames from a plurality of video streams generated from a plurality of cameras, each camera associated with a respective small model, each small model deployed in the edge device, sampling the plurality of frames to define sampled frames, performing inference to the sampled frames by using a big model, the big model shared by all of the plurality of cameras and deployed in a cloud or cloud edge, using the big model to generate labels for each of the sampled frames to generate training data, and training each of the small models with the training data to generate updated small models on the edge device.
1. A method for employing a semi-supervised learning approach to improve accuracy of a small model on an edge device, the method comprising:
collecting a plurality of frames from a plurality of video streams generated from a plurality of cameras, each camera associated with a respective small model, each small model deployed in the edge device;
sampling the plurality of frames to define sampled frames, including applying a mean Average Precision (mAP) policy that calculates a mAP for each frame based on a detection result from the small model and a ground truth using a detection result from the big model;
performing inference to the sampled frames by using a big model, the big model shared by all of the plurality of cameras and deployed in a cloud or cloud edge;
using the big model to generate labels for each of the sampled frames to generate training data; and
training each of the small models with the training data to generate updated small models on the edge device using a frame with a lowest mAP in a range.
2. The method of claim 1 , wherein sampling further includes applying a stride policy, the stride policy involves only using one frame from a group of frames in a range.
3. The method of claim 1 , wherein the training data is reduced by sampling further includes applying a confidence policy.
4. The method of claim 3 , wherein the confidence policy involves using a confidence value of a detection result from the small model to find a frame with a worst detection result.
5. The method of claim 3 , wherein a frame with a lowest confidence value is used in the training data.
6. A non-transitory computer-readable storage medium comprising a computer-readable program for employing a semi-supervised learning approach to improve accuracy of a small model on an edge device, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:
collecting a plurality of frames from a plurality of video streams generated from a plurality of cameras, each camera associated with a respective small model, each small model deployed in the edge device;
sampling the plurality of frames to define sampled frames, including applying a mean Average Precision (mAP) policy that calculates a mAP for each frame based on a detection result from the small model and a ground truth using a detection result from the big model;
performing inference to the sampled frames by using a big model, the big model shared by all of the plurality of cameras and deployed in a cloud or cloud edge;
using the big model to generate labels for each of the sampled frames to generate training data; and
training each of the small models with the training data to generate updated small models on the edge device using a frame with a lowest mAP in a range.
7. The non-transitory computer-readable storage medium of claim 6 , wherein sampling includes applying a stride policy, the stride policy involves only using one frame from a group of frames in a range.
8. The non-transitory computer-readable storage medium of claim 6 , wherein sampling further includes applying a confidence policy.
9. The non-transitory computer-readable storage medium of claim 8 , wherein the confidence policy involves using a confidence value of a detection result from the small model to find a frame with a worst detection result.
10. The non-transitory computer-readable storage medium of claim 8 , wherein a frame with a lowest confidence value is used in the training data.
11. A system for employing a semi-supervised learning approach to improve accuracy of a small model on an edge device, the system comprising:
a memory; and
one or more processors in communication with the memory configured to:
collect a plurality of frames from a plurality of video streams generated from a plurality of cameras, each camera associated with a respective small model, each small model deployed in the edge device;
sample the plurality of frames to define sampled frames, including a mean Average Precision (mAP) policy that calculates a mAP for each frame based on a detection result from the small model and a ground truth using a detection result from the big model;
perform inference to the sampled frames by using a big model, the big model shared by all of the plurality of cameras and deployed in a cloud or cloud edge;
use the big model to generate labels for each of the sampled frames to generate training data; and
train each of the small models with the training data to generate updated small models on the edge device using a frame with a lowest mAP in a range.
12. The system of claim 11 , wherein sampling further includes applying a stride policy, the stride policy involves only using one frame from a group of frames in a range.
13. The system of claim 11 , wherein sampling further includes applying a confidence policy.
14. The system of claim 13 , wherein the confidence policy involves using a confidence value of a detection result from the small model to find a frame with a worst detection result.