Research map / Vision-language and multimodal / CLIP and zero-shot

ProCC: Progressive Cross-Primitive Compatibility for Open-World Compositional Zero-Shot Learning

AAAI 2024 · Vision-language and multimodal › CLIP and zero-shot · approach: Attribute-based zero-shot learning · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Appendix of ProCC, a progressive, cross-primitive compatibility approach for open-world compositional zero-shot learning that predicts object and state primitives with conditioned classifiers trained in three stages.

Problem

Open-world compositional zero-shot learning must recognize novel state-object compositions from an output space containing all possible compositions with no priors, which is very large and contains many infeasible pairs.

Approach

Progressive three-stage training that first learns object classifiers, then state classifiers conditioned on object predictions via a cross-primitive compatibility module with a learnable cross-primitive memory, and finally finetunes all components jointly.

Key results

• On C-GQA in the OW-CZSL setting with the backbone updated, Ours_up reaches HM 5.3 and AUC 0.91 on test, above the best baseline Co-CGE_up (HM 4.8, AUC 0.78).
• On UT-Zappos (updated backbone) in OW-CZSL, Ours_up reports test HM 51.5 vs KGSP_up 52.1, so it is not best there; its test AUC is 27.9 vs 26.5 for KGSP_up.
• Ablations show 4 MLP layers and feature-dimension-sized CPM kernel (1/2 fd) give the best trade-off; training time of ProCC (4.7 on CGQA) is comparable to KGSP and lower than CANet.

Measured effects reported in the paper

Surprising findings

• Classifying states is substantially harder than classifying objects, and pre-trained condition information mainly helps state prediction.
• Joint training of the whole network (without progressive stages) causes the state and object conditioning to be noisy, motivating the progressive design.

Limitations

• Relies only on visual information; visually similar compositions are confused.
• Training is multi-stage, and joint training without progressive stages degrades performance.
• Performance on UT-Zappos is not uniformly best against KGSP.

Experimental setup

Datasets: UT-Zappos MIT-States C-GQA

Benchmarks: UT-Zappos MIT-States C-GQA

Metrics: seen accuracy unseen accuracy harmonic mean AUC

Baselines: CompCos Co-CGE CGE KGSP CANet

Builds on: KGSP CompCos Co-CGE CANet

Closest papers by meaning

• Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning (CVPR 2024 · similarity 0.89)
• Beyond Seen Primitive Concepts and Attribute-Object Compositional Learning (CVPR 2024 · similarity 0.88)
• Retrieval-Augmented Primitive Representations for Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.88)
• Beyond Image Classification: A Video Benchmark and Dual-Branch Hybrid Discrimination Framework for Compositional Zero-Shot Learning (CVPR 2025 · similarity 0.88)
• Compositional Few-Shot Class-Incremental Learning (ICML 2024 · similarity 0.87)
• Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning (AAAI 2024 · similarity 0.87)