AI Paper Atlas

Research map / Detection and segmentation / Weakly supervised detection and segmentation

CA-SSL: Class-Agnostic Semi-Supervised Learning for Detection and Segmentation

Lu Qi, Jason Kuen, Zhe Lin, Jiuxiang Gu, Fengyun Rao, Dian Li, Weidong Guo, Zhen Wen, Ming-Hsuan Yang, Jiaya Jia

ECCV 2022 · Detection and segmentation › Weakly supervised detection and segmentation · approach: Pseudo-labeling for semi-supervised detection · 13 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use class-agnostic rather than class-specific pseudo labels in warmup (ablation)CA-SSL warmup (objects) · COCO val2017 (tiny)AP det: 47.1 -> 48.1Improved
use pseudo labels rather than class-agnostic ground-truth labels in warmup (ablation)CA-SSL warmup (objects) · COCO val2017 (tiny)AP det: 46.9 -> 48.1Improved
add warmup on class-agnostic pseudo labels over ImageNet pretraining (ablation)Swin-T CondInst · COCO val2017 (tiny)AP det: 46.8 -> 48.1Improved
switch from box objects to mask entities in warmup labels (ablation)CA-SSL warmup · COCO val2017 panoptic (tiny)PQ: 42.5 -> 43.7Improved
switch from box objects to mask entities in warmup labels (ablation)CA-SSL warmup · COCO val2017 detection (tiny)AP det: 48.1 -> 48.2No clear effect
increase warmup epochs from 12 to 48 (ablation)CA-SSL warmup · COCO val2017 entity APAP^e: 38.4 -> 40.2Improved
increase warmup and finetuning from 12 epochs to 60 and 36 (ablation)CA-SSL · COCO val2017 AP detAP det: 45.3 -> 48.2Improved
use a stronger pseudo labeler (entity AP 38.0 to 42.1) (ablation)CA-SSL warmup · COCO val2017 warmed-up entity APAP^e: 36.1 -> 40.2Improved
add strong augmentation in both warmup and finetuning (ablation)CA-SSL · COCO val2017 AP detAP det: 48.2 -> 49.1Improved
scale unlabeled data from 123K to 3.66M images (ablation)CA-SSL · COCO val2017 AP detAP det: 48.2 -> 50.6Improved
initialize finetuning with warmup backbone (ablation)Swin-T-FCOS · COCO val2017 AP detAP det: 46.8 -> 50.6Improved
also initialize neck, head and classifier from warmup model (ablation)backbone-initialized model · COCO val2017 AP detAP det: 50.6 -> 51.5Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy