CoChat

Unsupervised Deep Video Hashing via Balanced Code for Large-Scale Video Retrieval

Open access
Wu, Gengshen, Han, Jungong, Guo, Yuchen, Liu, Li, Ding, Guiguang, Ni, Qiang, Shao, Ling · 2019 · IEEE Transactions on Image Processing

The DOI matches its Crossref record; 81 of the 96 references it cites were found in a registry or an index; 6 of the 38 sources its claims cite were checked for support.

Abstract

This paper proposes a deep hashing framework, namely Unsupervised Deep Video Hashing (UDVH), for largescale video similarity search with the aim to learn compact yet effective binary codes. Our UDVH produces the hash codes in a self-taught manner by jointly integrating discriminative video representation with optimal code learning, where an efficient alternating approach is adopted to optimize the objective function. The key differences from most existing video hashing methods lie in 1) UDVH is an unsupervised hashing method that generates hash codes by cooperatively utilizing feature clustering and a specifically-designed binarization with the original neighborhood structure preserved in the binary space; 2) a specific rotation is developed and applied onto video features such that the variance of each dimension can be balanced, thus facilitating the subsequent quantization step. Extensive experiments performed on three popular video datasets show that UDVH is overwhelmingly better than the state-of-the-arts in terms of various evaluation metrics, which makes it practical in real-world applications.

Overview

CoChat Summary

TL;DR

The paper proposes Unsupervised Deep Video Hashing (UDVH), a self-taught deep hashing framework that jointly learns video representations and balanced binary codes via feature clustering, CCA, and a balanced rotation. On FCVID, YFCC, and ActivityNet, UDVH-TSN outperforms state-of-the-art video hashing baselines in retrieval accuracy and training efficiency.

Key findings

  • UDVH-TSN achieves mAP@20 of 50.4% on FCVID at 128 bits, 4.8% higher than UDVH-LSTM (45.6%) and 8.9% over SSTH; on YFCC it reaches 29.8% vs. 26.1% and 24%; on ActivityNet it reaches 26.2% vs. 20.7% and 17.8%.
  • Precision@100 at 64 bits is 51% on FCVID (at least about 3.6% higher than comparable baselines), 30.2% on YFCC vs. SSTH’s 26.3%, and 26.1% on ActivityNet vs. SSTH’s 20.4%.
  • The balanced rotation (CCA-BR) outperforms CCA-ITQ and PCA-BR, producing flatter per-dimension variance; cluster number around 400 gives best mAP@20 at 128 bits (FCVID 0.504, YFCC 0.293, ActivityNet 0.262).
  • Training is efficient: UDVH-TSN takes about 0.74–0.81 hours per loop and less than 4.5 hours total, versus over 8 hours for SSTH; balanced rotation costs only seconds.

Methods

  • UDVH uses k-means on 10,000 sampled FC7 1024-d video features to generate C pseudo-labels, CCA to project features to k dimensions, and a balanced rotation solved as an orthogonality-constrained l2,1-norm maximization; it alternates updating R, B = sign(H × R), and deep parameters Θ via SGD minimizing ||F(Z; Θ) − B||_F^2, with t = 3–5 loops.
  • The deep architecture uses modified Temporal Segment Networks (TSNs) with spatial and temporal ConvNets, fused global-pool features, and two fully connected layers (FC7 and FC8); 20 frames are uniformly sampled per video.
  • Experiments use FCVID (91,223 videos, 239 categories; 45,611 train), YFCC (0.7M processed videos; 0.6M unlabeled train, 0.1M labeled test), and ActivityNet (~20,000 clips, 200 categories; ~10,000 train/test; 1,000 queries), comparing against AGH, SubMod, ITQ, SP, MFH, DH, SSTH, and UDVH-LSTM with mAP@K, PR, and Precision@100 at 16/32/64/128 bits.

Limitations

  • The paper states ActivityNet gives the worst performance due to many untrimmed videos and large intra-class variance, with improvement there described as “a bit limited”; YFCC performance drops because many videos are low-quality amateur mobile clips; loss-function effects are not obvious (less than 3% improvement).
  • Evident from the text: cluster number C, iteration count t, and other hyperparameters are chosen empirically, and the alternating optimization may only find a local optimum.
  • Evident from the text: evaluation is limited to three video datasets and a selected set of baselines, and the method is unsupervised, relying on pseudo-labels rather than manual labels.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tip.2018.2882155

Cites

81 of 81 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 8 Not valid · 22 Not checked · 75 Own results and interpretations

53 more claims

Create a free account to see all 105 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 42
  • Probably not a substantial claim. 7

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 81 references

Research works printed but not linked: 3

  1. over, 2014, enot, “TRECVID 2014—An overview of the goals, tasks, data, evaluation mechanisms and metrics, Proc TRECVID, 52
  2. xu, 2013, Harmonious hashing, Proc IJCAI, 1820
  3. liu, 2011, Hashing with graphs, Proc ICML, 1

Cited by

Items 101–149 of 149

Metadata

Authors
Wu, G., Han, J., Guo, Y., Liu, L., Ding, G., Ni, Q., & Shao, L.
Affiliations
  • Lancaster University (GB)
  • Tsinghua University (CN)
  • Inception Institute of Artificial Intelligence (AE)
Venue
IEEE Transactions on Image Processing
Published
2019-04
Volume
28
Issue
4
Pages
1993-2007
ISSN
  • 1057-7149 (ISSN-L)
  • 1941-0042
OpenAlex
W2900790473
Cited by
148
References
96
Citations by year
2026: 2 · 2025: 9 · 2024: 9 · 2023: 21 · 2022: 24 · 2021: 22 · 2020: 29 · 2019: 32
FWCI
10.38
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Image Processing (publishedVersion)
  • PubMed (publishedVersion)
  • Lancaster EPrints (Lancaster University) (acceptedVersion)
  • Warwick Research Archive Portal (University of Warwick) (submittedVersion)
Funders
  • Royal Society (IE150997)
  • Shenzhen Government (GJHZ20180419190732022)
  • National Natural Science Foundation of China (61773301, 61571269)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Retrieval and Classification Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Analysis and Summarization
Keywords
video hashing, binary codes, feature clustering, binarization, quantization, alternating optimization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.