Post Process Weighing Algorithm

In order to post process multiple detections in a single frame to a usable single-number metric, we first use a post-processing weighing algorithm. After running the algorithm we send the value to the EMA tracking algorithm. This page goes over the Post-Processing Weighing Algorithm

Sigmoid-Logit Weighing

The output from the ML model will be in the format of:

[
[x1, xy1, x2, y2, conf, class], //what each index pertains to
[0.1, 0.15, 0.35, 0.45, 0.88, 1.0], //example values
...
]

The tensor of detections is then run through the following equations:

Where:

• X_c is the logit function modelled portion of the equation. This relates to the linear portion of the output (post simgoid function).
• If there is only a X_c element of the output tensor, then the post-processed score will be a 1:1 linear representation of the input (e.g. no change to the value)
• If there are elements other than X_c, then the final output will not be 1:1 to the input
• Z is the combination of the Linear portion (X_c) and the remaining detection scores C_i for i = 1 to N
• ϕ is the sigmoid function (𝜎) applied to Z
• 𝛾 is a constant value >= -1 used to weigh the remaining values with respect to the maximum detection value

Visualization

Below are a few plots of the post-processing weighing algorithm with varying values of 𝛾 and X_c. One value C_i is added to the detections where C_i = X_c - 0.05

Scenario 1: 𝛾 = -0.5:

Scenario 2: 𝛾 = 0.0:

Scenario 3: 𝛾 = 1.0:

Scenario 4: 𝛾 = 4.0: