Open cczarnuch opened 3 years ago
For this project I think we should use Deep SORT. For detection it uses YOLOv3 I believe and stored appearance information for tracking so if there is some level of occlusion the object won't be re-detected making our results more accurate. A sample git repo we can use is here.
A recent benchmarking paper was released for optimizers. The latest revision was published on the 5th, so we should definitely reference it for our model tuning. https://arxiv.org/abs/2007.01547
From what I looked at, it seems like Adam still wins out comparatively for image recognition on CNNs
Here's the most useful graphic though:
You can compare an optimizer with one shot (the default parameters) vs. models trained with a small budget of time/data. Diagonal shows how the model improves with training on average with that optimizer
Create a solution to provide object tracking for each of the following categories: