All Stories
Anchor Boxes in Object Detection
The one-object-per-cell limit anchor boxes remove, the label vector written out in full, and why k-means on the training set beats hand-picked shapes.
01 Jun 2023 · 4 min read
Non-Max Suppression
How detectors collapse dozens of overlapping boxes for one object into a single prediction, the exact algorithm, and the failure it causes in crowds.
30 May 2023 · 3 min read
Intersection over Union
The overlap metric every detector is scored on, worked with real coordinates, and why two boxes that look close can still fail the 0.5 threshold.
28 May 2023 · 4 min read
The State of Computer Vision
The data-versus-hand-engineering spectrum, why detection sits further along it than classification, and which benchmark tricks not to ship.
15 Apr 2023 · 3 min read
Data Augmentation in Computer Vision
The standard augmentations and what invariance each one encodes, PCA colour augmentation written out, and why the wrong transform quietly costs accuracy.
14 Apr 2023 · 4 min read
Transfer Learning
How much of a pre-trained network to freeze given your dataset size, why features stop transferring as you go deeper, and the caching trick that makes it fast.
13 Apr 2023 · 4 min read