Research map / Image generation / Autoregressive and latent generation
Distributed Manifold Hashing for Image Set Classification and Retrieval
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.
Proposes Distributed Manifold Hashing (DMH), which learns binary hash codes for image sets stored across a network of nodes using Riemannian (SPD/Grassmann) kernels and ADMM-based parallel optimization with neighbor consistency constraints.
Problem
Image set classification and retrieval when image sets are distributed across multiple nodes, with efficiency (storage and computation) for large-scale data.
Approach
Models distributed image sets as a connected graph, represents each image set with a Riemannian manifold kernel, and learns per-set hash codes jointly with a classifier via distributed ADMM where global variables are constrained to match only among neighboring nodes.
Key results
- DMH-SPD reaches 98.72% accuracy on BBT, 98.03% on PB and 76.48% on YTC, improving the best baselines by 0.51%, 5.31% and 1.21%.
- Hash codes require 64x less storage than continuous 32-d features and under 100 KB total across the three datasets.
- Hashing distance computation is about 8x faster than GDA/CDL and about 700x faster than PML on BBT; 8-node parallel training is roughly 10x faster than single-node training on PB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMH-SPD instead of the strongest baseline | BBT image set classification · BBT | Accuracy: 98.21 -> 98.72; 0.51% | Helped |
| Use DMH-SPD instead of the strongest baseline | PB image set classification · PB | Accuracy: 92.72 -> 98.03; 5.31% | Helped |
| Use DMH-SPD instead of the strongest baseline | YTC image set classification · YTC | Accuracy: 75.27 -> 76.48; 1.21% | Helped |
| Use DMH instead of the strongest baseline | BBT image set retrieval · BBT | mAP: 96.00 -> 99.14 | Helped |
| Use DMH-SPD instead of the strongest baseline | PB image set retrieval · PB | mAP: 91.40 -> 96.69 | Helped |
| Use DMH-SPD instead of the strongest baseline | YTC image set retrieval · YTC | mAP: 75.65 -> 79.37 | Helped |
| Add the binary hash-code constraint (ablation) | DMH-SPD · BBT | Accuracy: 98.70 -> 98.72 | No clear effect |
| Add the binary hash-code constraint (ablation) | DMH-SPD · PB | Accuracy: 97.86 -> 98.03 | No clear effect |
| Add the binary hash-code constraint (ablation) | DMH-SPD · YTC | Accuracy: 68.68 -> 76.48 | Helped |
| Add the binary hash-code constraint (ablation) | DMH-GM · BBT | Accuracy: 91.67 -> 96.38 | Helped |
| Add the binary hash-code constraint (ablation) | DMH-GM · PB | Accuracy: 83.52 -> 96.19 | Helped |
| Add the binary hash-code constraint (ablation) | DMH-GM · YTC | Accuracy: 53.52 -> 73.39 | Helped |
Surprising findings
- Binary hash codes sometimes match or exceed their continuous variant (e.g., DMH-SPD 98.72% vs 98.70% on BBT accuracy), and DMH-GM mAP on BBT (99.14) exceeds DMH-SPD (97.30).
Limitations
- Baselines such as AHISD, CHISD and LEML could not be run on large datasets (exceeded one week), so their results are missing.
- Classification at test time uses a local hash function from one randomly selected node, and accuracy slightly drops as node count grows.
- Evaluated only on three face/video-frame image set datasets with a single fixed deep feature extractor.
Experimental setup
Datasets: BBT PB YTC
Benchmarks: BBT PB YTC
Metrics: accuracy mAP storage running time
Baselines: MSM DCC GDA CDL PML ITQ SDH
Builds on: Supervised Discrete Hashing (SDH) Covariance Discriminative Learning (CDL) Grassmann Discriminant Analysis (GDA) Projection Metric Learning (PML) Discrete cyclic coordinate descent
Closest papers by meaning
- An ensemble diversity approach to supervised binary hashing (NeurIPS 2016 · similarity 0.84)
- Large Graph Hashing with Spectral Rotation (AAAI 2017 · similarity 0.83)
- Deep Supervised Hashing for Fast Image Retrieval (CVPR 2016 · similarity 0.83)
- S2JSD-LSH: A Locality-Sensitive Hashing Schema for Probability Distributions (AAAI 2017 · similarity 0.83)
- Dual Manifold Regularization Steered Robust Representation Learning for Point Cloud Analysis (AAAI 2025 · similarity 0.83)
- Deep Graph Online Hashing for Multi-Label Image Retrieval (AAAI 2025 · similarity 0.82)
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.