Skip to content

GitLab

Commit 6cf9a28c authored by Scott Main's avatar Scott Main Committed by TF Object Detection Team
Browse files

Add link to notebook that generates anchor box ratios with k-means clustering 

PiperOrigin-RevId: 407366384
parent fdd0a8a3
Hide whitespace changes
Inline Side-by-side
Showing with 7 additions and 0 deletions
research/object_detection/g3doc/configuring_jobs.md
View file @ 6cf9a28c
...@@ -104,6 +104,13 @@ increase computation costs. Whereas generating fewer anchors that have a higher ...@@ -104,6 +104,13 @@ increase computation costs. Whereas generating fewer anchors that have a higher
chance to overlap with ground truth will both improve accuracy and reduce chance to overlap with ground truth will both improve accuracy and reduce
computation costs. computation costs.
And although you can manually select values for both scale and aspect ratios
that work well for your dataset, there are programmatic techniques you can use
instead. One such strategy to determine the ideal aspect ratios is to perform
k-means clustering of all the ground-truth bounding-box ratios, as shown in this
Colab notebook to [Generate SSD anchor box aspect ratios using k-means
clustering](https://colab.sandbox.google.com/github/tensorflow/models/blob/master/research/object_detection/colab_tutorials/generate_ssd_anchor_box_aspect_ratios_using_k_means_clustering.ipynb).
**Single Shot Detector (SSD) full model:** **Single Shot Detector (SSD) full model:**
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment