Build process creates training, validation and testing sets from database of clips.
Changed build process to create a dataset which works with frames or segments ( e.g. 45 frames).
Frames (important frames) are calculated by choosing frames with a mass ( the count of pixels that have been deemed object by track extraction) between the LQ and UQ.
Frames are also checked to see if they are noisy frames.
This attempts to remove the following tyeps of noisy frames:
Frames where it has bad tracking e.g. in the middle of an animal
Frames where only part of the animal is shown (maybe leaving the cameras field of vision)
Frames where nothing is there but noise
Segments are calculated either by:
Choosing random permutations of the important frames.
Number of Segments
(# of important frames - segment duration) // 9 segments are selected
or by choosing segment duration consecutive frames whose mass is above a certain amount
Number of Segments
up to (# of frames - segment duration) // segment-frame-spacing
split datasets by camera and location ( too try and remove any bias that may occur from using a camera in multiple sets).
Updated Cumulative Distribution Function
so that segments/frames will only be used once per epoch ( round of training).
Added ability to weight probabilities of each label ( e.g. 80 bird per 20 false positives)
Removed uneeded code for heavy bins
Some labels have low amounts of data so a single camera is split into 2 e.g. Wallabies and Leoparidaes
Added mapped_labels which lets you group multiple labels into a single label. e.g. possum, hedgehog and rabbit into the labels pests
Build process creates training, validation and testing sets from database of clips. Changed build process to create a dataset which works with frames or segments ( e.g. 45 frames).
Frames (important frames) are calculated by choosing frames with a mass ( the count of pixels that have been deemed object by track extraction) between the LQ and UQ.
Frames are also checked to see if they are noisy frames. This attempts to remove the following tyeps of noisy frames:
Segments are calculated either by: Choosing random permutations of the important frames.
or by choosing segment duration consecutive frames whose mass is above a certain amount
split datasets by camera and location ( too try and remove any bias that may occur from using a camera in multiple sets).
Updated Cumulative Distribution Function
Removed uneeded code for heavy bins
Some labels have low amounts of data so a single camera is split into 2 e.g. Wallabies and Leoparidaes
Added mapped_labels which lets you group multiple labels into a single label. e.g. possum, hedgehog and rabbit into the labels pests