Neglected Free Lunch

Learning from the traces left during annotation

Neglected Free Lunch – Learning Image Classifiers Using Annotation Byproducts

Han et al. · ICCV 2023 · CC BY 4.0

Contribution

Contributed major engineering work to the ImageNet and COCO annotation tools and co-authored the paper. Both official repositories credit Dante (@1000ship) with most of the work in their respective repositories.

Research collaboration with NAVER AI Lab researchers. The paper lists Seonghyeok Chun's affiliation as Dante Company.

ImageNet · Contribution evidence ↗COCO · Contribution evidence ↗

During training, images, labels and annotation byproducts support classification plus auxiliary location regression. At inference, only the image is required. — Original explanatory diagram based on Han et al. (ICCV 2023); not a paper figure or experiment screenshot.

Research question

Keeping only images and labels discards behavioral information generated during annotation.

Method

LUAB uses annotation byproducts, including clicks, as weak location signals while jointly learning image classification and auxiliary location regression. Byproducts and the regression head are not required at inference.

Results reported in the paper

Model / Dataset	Metric	Baseline → LUAB	Source
ViT-B / ImageNet1K	Classification accuracy	81.6 → 82.5 (+0.9%p)	Table 1 ↗
ResNet50 / COCO	Multi-label classification mAP	73 → 74.2 (+1.2%p)	Table 6 ↗

Paper and public resources

Authors: Dongyoon Han, Junsuk Choe, Seonghyeok Chun, John Joon Young Chung, Minsuk Chang, Sangdoo Yun, Jean Y. Song, Seong Joon Oh.

Seonghyeok Chun — Dante Company

Paper · 2023 | Public tool commit · May 3, 2022

ICCV 2023 ↗