Redefining the One-Shot Object Detector for a Two-Class Problem
Type
Work project
From
UncannyVision / Eagle Eye Networks
Headline result
AP head 88.5% → 91%, AP body 87.5% → 91.9%
Topics
object detection
Write-up
Description: One shot object detectors face the anchor problems when the dataset is biased and high variance in the box sizes. To tackle that issue, I developed a hierarchial clustering approach to adjust the anchors height and width viz. Anchor Search algorithm for assigning and learning optimal anchors for two different classes individually. This was developed by taking into consideration the two classes i.e person body and head which somewhat vary a lot in the size and aspect ratio. Also entangled the training of head and body of person together instead of treating them as two different classes 3% mAP gains were observed after final bench-marking the improvements.
Results on the final improvements.
| Detector Model | AP Head | AP Body |
|---|---|---|
| In-house Vanilla Detector | 88.5% | 87.5% |
| In-house Vanilla Detector + Anchor Search | 90.02% | 90.11% |
| In-house Vanilla Detector + Anchor Search + Entangle | 91% | 91.9% |