AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

Revisiting Image Aesthetic Assessment via Self-Supervised Feature Learning

Kekai Sheng, Weiming Dong, Menglei Chai, Guohui Wang, Peng Zhou, Feiyue Huang, Bao-Gang Hu, Rongrong Ji, Chongyang Ma

AAAI 2020 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning · 43 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
Jointly train degradation identification and triplet loss instead of using either alone (ablation)Aesthetics-aware pretext task · AVA and AADBaccuracy: ∼ 0.3% on the AVA dataset and ∼ 0.7% on the AADB datasetImproved
Add triplet loss to degradation identification training (ablation)Aesthetics-aware pretext task · AVA and AADBaccuracy: ∼ 0.3% on the AVA dataset and ∼ 0.7% on the AADB datasetImproved
Use triplet loss alone rather than the two-stage learning scheme (ablation)Aesthetics-aware pretext task · AVA and AADBtraining stabilityHurt
Add JPEG compression and Gaussian noise operations (ablation)Proposed self-supervised pretraining · AVA and AADBaccuracy: ∼ 0.1% on AVA; ∼ 0.3% on AADBImproved
Add exposure manipulation (ablation)Proposed self-supervised pretraining · AVA and AADBaccuracy: ∼ 0.3% on AVA; ∼ 0.4% on AADBImproved
Add downsampling, quantization, and pixelation operations (ablation)Proposed self-supervised pretraining · AVA and AADBaccuracy: ∼ 0.4% on AVA; 0.5 ∼ 0.8% on AADBImproved
Add Gaussian blur (ablation)Proposed self-supervised pretraining · AVA and AADBaccuracy: ∼ 0.1% on AVA; ∼ 0.2% on AADBImproved
Add rotation operation (ablation)Proposed self-supervised pretraining · AVA and AADBaccuracy: 0.3 ∼ 0.5% on AVA; 0.4 ∼ 0.6% on AADBImproved
Add patch shuffle and mixup operations (ablation)Proposed self-supervised pretraining · AVA and AADBaccuracy: ∼ 0.2% on AVA; 0.3 ∼ 0.4% on AADBImproved
Add entropy-based weights to patches (ablation)Proposed self-supervised pretraining · Task adaptation benchmarksaccuracy: at least 0.5%Improved
Use the proposed pretext task rather than ImageNet-label pretrainingImage aesthetic assessment · AVA, AADB, and CUHK-PQaccuracyMixed

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