AI Paper Atlas

Research map / Detection and segmentation / CNN features

Dynamic Low-Resolution Distillation for Cost-Efficient End-to-End Text Spotting

Ying Chen, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, Xi Li

ECCV 2022 · Detection and segmentation › CNN features · approach: Arbitrary-shape scene text detection · 7 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
add sequential knowledge distillation at half resolution (ablation)Vanilla Multi-Scale low-res · Total-TextEnd-to-End None: 55.4 -> 59.6Improved
add sequence-level beam search loss L_seq alone (ablation)SKD-only · Total-TextEnd-to-End None: 55.4 -> 58.4Improved
add all three SKD losses (ablation)SKD-only · Total-TextEnd-to-End None: 55.4 -> 59.6Improved
replace sequence-level distillation with logits-based distillation (ablation)SKD · Total-TextEnd-to-End None: 59.6 -> 58.7Hurt
add detection distillation of Qi et al. to SKD (ablation)SKD · Total-TextEnd-to-End None: 59.6 -> 59.8No clear effect
add dynamic resolution selector to SKD (full DLD, gamma 0.1)SKD-only · Total-TextEnd-to-End None: 59.6 -> 63.9Improved
use DLD instead of high-res vanilla inputVanilla Multi-Scale · IC15End-to-End General: 69.5 -> 70.9Improved
use larger RoI scale 16x64 in teacher and student (ablation)DLD · Total-TextEnd-to-End None: 63.9 -> 64.8Improved
align smaller student RoI with deconv instead of using larger RoI (ablation)DLD · Total-TextEnd-to-End None: 64.8 -> 64.5Hurt
use larger candidate scales {0.8,0.7,0.6} instead of {0.5,0.4,0.3} (ablation)DLD · Total-TextEnd-to-End None: 58.9 -> 63.6Improved
use six candidate scales instead of three larger ones (ablation)DLD · Total-TextEnd-to-End None: 63.6 -> 63.9Improved
increase gamma from 0.1 to 0.5 (ablation)DLD · Total-TextEnd-to-End None: 63.9 -> 56.0Hurt

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