Source-linked AI summary
A Deeper Look at Dataset Bias
Tatiana Tommasi, Novi Patricia, Barbara Caputo, Tinne Tuytelaars
TL;DR
The paper asks whether DeCAF features overcome dataset bias across differing image collections, where prior evidence was limited. It evaluates DeCAF and debiasing approaches on a broad cross-dataset testbed and finds that bias persists, existing debiasing methods generally fail, and some adaptive strategies are highly effective.
Problem
Dataset bias limits generalization when training and test collections differ, and prior analyses did not establish whether DeCAF resolves this problem in unconstrained cross-dataset settings.
Method
The paper performs extensive cross-dataset analyses of DeCAF features and evaluates existing debiasing and domain-adaptation methods across a twelve-dataset testbed.
Results
Dataset bias persists with DeCAF, existing ad-hoc debiasing and domain-adaptation methods do not help, while some previously discarded adaptive strategies appear extremely effective.
Takeaways & Limitations
Changing feature representation does not intrinsically remove negative dataset bias, leaving cross-database generalization open for new research directions.
Takeaways & Limitations
The name-the-dataset analysis cannot reveal how classes relate across collections or how a model for one class generalizes to others.
Abstract
from arXiv · showhide
The presence of a bias in each image data collection has recently attracted a lot of attention in the computer vision community showing the limits in generalization of any learning method trained on a specific dataset. At the same time, with the rapid development of deep learning architectures, the activation values of Convolutional Neural Networks (CNN) are emerging as reliable and robust image descriptors. In this paper we propose to verify the potential of the DeCAF features when facing the dataset bias problem. We conduct a series of analyses looking at how existing datasets differ among each other and verifying the performance of existing debiasing methods under different representations. We learn important lessons on which part of the dataset bias problem can be considered solved and which open questions still need to be tackled.
1 Introduction
The paper examines whether DeCAF features overcome dataset bias, extending evaluation beyond controlled domain shifts and supervised transfer settings. Its broad experiments find that dataset bias persists, existing debiasing methods generally fail, and some adaptive strategies are highly effective.
- Motivation: The study tests whether DeCAF features resolve dataset bias when training and testing collections differ without target labels.Prior analyses largely considered controlled visual shifts, labelled target samples, or CNN fine-tuning.
- Contributions: The authors evaluate DeCAF on a twelve-dataset cross-dataset testbed and introduce a measure combining in-dataset performance with cross-dataset performance loss.The measure addresses the limitation of percentage drop alone, which ignores the source-dataset performance level.
- Scope: The analysis covers cross-dataset generalization, undoing dataset bias, and the potential usefulness of domain adaptation algorithms.These analyses are conducted under different image representations, with particular focus on DeCAF features.
- Main findings: The negative bias persists, existing ad-hoc debiasing algorithms do not help, and some previously discarded adaptive strategies appear extremely effective.The findings characterize dataset bias as an open problem requiring directions that accommodate both deep features and cross-database generalization difficulties.
2 What is Dataset Bias?
Dataset bias arises because finite, task-specific collections capture limited and distinctive aspects of the visual world. The paper frames generalization across collections through domain shift, sample selection bias, and differences in marginal and conditional distributions.
- Sources of bias: Finite, task-specific image collections cover only parts of the visual world, making their dataset-specific bias recognizable.Capture, category, and other collection-specific biases contribute to these differences.
- Why bias matters: Dataset bias can produce false conclusions because datasets created to evaluate learning models are limited in scope.The central question is how to generalize to unseen samples when training and test collections differ.
- Formal framing: A domain is defined by marginal and conditional label-related distributions, while sample selection bias primarily reflects marginal-distribution differences.When collections share categories, representation choices can remain a major source of bias.
- Generalization: Unseen cross-domain error depends on training error, domain divergence, and differences between the collections’ conditional distributions.Cross-dataset performance drop alone omits the importance of source-dataset training error.
- Related methods: Domain adaptation methods address related dataset-bias settings, but prior evaluations often use few samples or do not suit large datasets.The paper motivates reassessing the problem given current collection sizes and CNN feature performance.
3 Evaluation Protocol
The evaluation uses twelve datasets, multiple feature representations, and both in-dataset and cross-dataset testing. It supplements percentage performance drop with a bounded measure that incorporates the source performance level.
- Datasets: The testbed contains sparse and dense setups: the sparse setup spans twelve datasets, while the dense setup contains 40 shared classes from four datasets.The sparse setup includes differing numbers of aligned classes across collections.
- Features: The released testbed provides BOWsift and DeCAF6/DeCAF7 representations for cross-dataset experiments.DeCAF descriptors are 4096-neuron activation vectors from the sixth and seventh CNN layers.
- Features: Features are L2-normalized, while BOWsift receives z-score normalization for domain-adaptation tests and serves mainly as a reference baseline.The experiments focus primarily on DeCAF representations.
- Evaluation: In-dataset evaluation trains and tests within one collection, whereas cross-dataset evaluation trains and tests on different collections using Self and Mean Other performance.Self denotes in-dataset performance; Mean Other averages performance across multiple test collections.
- Measures: 75% percentage drop can correspond to different outcomes because the metric ignores the Self performance level.The proposed CD measure uses Self − Mean Other and rescales it with a sigmoid between 0 and 1.
4 Studying the Sparse set
The sparse-set experiments show that DeCAF captures dataset-specific content effectively but does not eliminate dataset bias. Cross-dataset behavior varies by object class, representation, and dataset, while existing debiasing can underperform pooling all source data.
- Name the Dataset: DeCAF improves dataset recognition for A-Yahoo, MSRCORID, Pascal VOC07, and SUN, while reducing confusion among Bing, Caltech256, and Imagenet.ETH80, Office, and RGB-D remain easy to distinguish because of their lab-oriented collection characteristics.
- Name the Dataset: Dataset recognition gives only a partial view because it does not show class relationships or how models trained for one dataset generalize to others.The experiments therefore extend to cross-dataset tests on shared car and cow categories.
- Cross-dataset generalization test: Scene-centric and object-centric datasets remain separated, while object-centric collections achieve higher cross-dataset results than their Self performance.This pattern is more evident with DeCAF than with BOWsift, partly because multi-label scene collections create training and testing confusion.
- Cross-dataset generalization test: DeCAF makes car transfer less dataset-specific but leaves a large in-dataset versus cross-dataset gap for cow, whose greater in-class variability makes recognition harder.Using a fixed test negative set increases cow cross-dataset results, indicating that negative-set bias persists across representations.
- Cross-dataset generalization test: DeCAF cannot fully solve dataset bias: it helps in some class-dependent cases, can worsen bias by capturing finer image details, and does not remove negative bias.The reported comparison includes car, ETH80, and MSRCORID as cases where DeCAF can help.
- Undoing the Dataset Bias: When comparing Unbias with pooling all source images, Unbias usually helps BOWsift but the pooled All SVM generally performs better with DeCAF7.Pooling improves coverage of object variability and therefore yields higher cross-dataset performance in the reported experiments.
5 Studying the Dense set
The dense cross-dataset experiments show that DeCAF improves dataset characterization but does not eliminate dataset bias. In noisy-source adaptation, existing subsetting and grouping methods fail to outperform using all source data, whereas self-labelling consistently improves results.
- 5.1 Name the Dataset: DeCAF7 outperforms DeCAF6 and BOWsift in the balanced 40-class name-the-dataset experiment.The dense setup uses Caltech256, Bing, SUN, and Imagenet, with five test samples per class and one to fifteen training samples per class.
- 5.1 Name the Dataset: DeCAF separates object- and scene-centric datasets, while making differences among object-centric collections more evident than BOWsift.SUN is scene-centric; Bing, Caltech256, and Imagenet are object-centric.
- 5.2 Cross-dataset generalization test: DeCAF7 produces asymmetric cross-dataset behavior, with Caltech256 and Imagenet generalizing more easily to each other than to SUN.The authors caution that DeCAF was defined over Imagenet samples, which may contribute to the asymmetry.
- 5.2 Cross-dataset generalization test: Dataset bias remains class dependent: DeCAF does not automatically solve it, and cross-dataset recognition varies substantially by object class.For Caltech256 training, motorcycle, car, people, spoon, umbrella, basketball-hoop, and laptop remain among classes with large cross-dataset differences.
- 5.4 Experiments: Landmark selection and reshape-based grouping do not outperform using all noisy source data, while self-labelling consistently improves the original results.The gain is especially significant when only a reduced number of training images per class is available.
- 5.3.1 Methods: With DeCAF features, self-labelling’s usual annotation-drift risk appears highly reduced across ten target-selection iterations compared with BOWsift.The procedure selects high-confidence target samples using classifier margins and repeats for ten iterations.
6 Conclusions
The paper conducts an extensive evaluation of dataset bias using DeCAF features, finding that their strong descriptive power does not generally resolve cross-dataset generalization and can make learning general information harder.
- The evaluation examines dataset bias through an extensive experimental protocol centered on DeCAF, a popular CNN-learned image descriptor.The study expands the scale and complexity of the evaluation to analyze different aspects of the problem.
- DeCAF does not generally solve dataset bias and can perform worse than less powerful BOWsift features in some class- and dataset-dependent cases.The negative bias also remains because changing feature representation does not intrinsically remove or alleviate it.
- Neither dataset-specific debiasing methods nor domain-adaptation algorithms work effectively over DeCAF, whereas self-labeling the test set significantly increases performance.The results suggest that DeCAF’s highly descriptive representations make learning general information across collections more difficult.