Skip to main content

[WIP] Semantic Segmentation from Dashcam

Ongoing notes, exploration, and development
Created on January 23|Last edited on March 26

Section 9




05101520Step0.40.50.60.70.80.9
Run set
2


Overview




Run set
2


First Pass: Increase Weight Decay, Decrease Learning Rate




All manual runs
107
First manual sweep
10


Hyperparameter Sweep Insights




Run set
398


Manual vs Automated Sweeps




Runs by sweep
398


Per-class accuracies




Run set
5


Resnet v. Alexnet




Run set
2


New Approaches to Semantic Segmentation

How can this model be improved?

  • data verification: figure out which dataset items feature humans, and try training/testing on those explicitly
  • metrics: use mean IoU across classes--could also split this up per class
  • SOTA version: the SOTA model for semantic segmentation is the HRNet. Some other models and code for a semantic segmentation benchmark may also be inspiring.

Literature review

Object-contextual representations

Input size for existing model is [720, 1280], while for HRNet the options are [512, 1024] and [1024, 2048]. Perhaps rescaling is sufficient.

Section 8




Run set
1