Unbiased Look at Dataset Bias

“Unbiased Look at Dataset Bias” is a paper by Antonio Torralba of the Massachusetts Institute of Technology and Alexei A. Efros of Carnegie Mellon University, presented at the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) in Colorado Springs in June 2011. It asked whether the datasets that object recognition research trained and tested on still represented the visual world they had been collected to sample, and answered with four experiments on popular datasets: a classifier that guesses which dataset an image came from, a train-on-one, test-on-another matrix for “car” and “person” models, a test of how each dataset’s negative examples compare with everyone else’s, and an estimate of how many examples from one dataset a training example from another is worth. It also named four kinds of bias and suggested how dataset builders might reduce three of them. This article covers the publication as read in the authors’ copy, whose page numbers (1521–1528) match the proceedings; section, table and figure numbers refer to that copy. The concept as it is understood today, including bias in datasets of people, is the subject of dataset bias.

Problem

Section 2 credits datasets with much of the recent progress in recognition, both as training data and as the means of comparing methods, and then lists what that success had cost (§2). One concern was “creeping overfitting”: methods tuned over years to one dataset’s idiosyncrasies, which the authors note PASCAL VOC had countered by changing its data regularly (2005–2007) and, from 2008, by withholding the test set and limiting evaluations, as did the Caltech Pedestrian benchmark. Another was the weight given to winning a competition whether or not the margin was statistically significant. Citing Everingham et al.’s use of the Friedman test with Nemenyi post hoc analysis, the authors say it found no significant difference among the eight top-ranked methods of the 2010 VOC competition, and they suggest treating benchmark numbers as a sanity check and a quick comparison with baselines rather than a race. The cited IJCV paper applies that test to the VOC2007 classification entries; the 2010 analysis is not in it. Both issues, the authors judged, were behavioral and would fade as the field adopted better benchmarking practice; benchmarks and leaderboards and statistical significance in evaluation cover them today.

The more fundamental question was whether the datasets measured the right thing. In machine learning, the authors write, the dataset is the world, whereas a computer vision dataset is supposed to represent the world, and several had instead become “closed worlds” of their own: the Corel world, the Caltech-101 world, the PASCAL VOC world (abstract and §2). Their history of datasets (§2.1) is a sequence of reactions to the previous generation’s biases: Lena, digitized in 1972 from a Playboy centerfold, against controlled laboratory images; COIL-100’s household objects on black backgrounds against model-based thinking; Corel and 15 Scenes against simple backgrounds; Caltech-101, mined with Google, against Corel’s professionalism; MSRC and LabelMe, with many objects per scene, against the single centred object; PASCAL VOC against lax training and testing standards; and Tiny Images, ImageNet and SUN09 against datasets too small for the real world. Each new dataset, they argue, rejected the old biases and acquired similar ones, because the field lacked a clear understanding of the types and sources of bias.

Contribution

• Name That Dataset as a diagnostic: if a classifier can tell which dataset an image came from, the datasets differ in ways a model can learn (§1).
• A cross-dataset generalization matrix for two classes and two tasks across six datasets, summarized per training set by its own-test score, its mean score on the others and the relative drop (§3.1, Table 1).
• A negative-set test that keeps each dataset’s positives and replaces its test negatives with negatives pooled from all six datasets (§3.2, Table 2).
• The “market value” of a training sample, the number of examples of one dataset that one example of another is worth, read from learning curves (§4, Table 3).
• Four named kinds of bias, selection, capture, category or label, and negative-set bias (§3.1), with suggestions for reducing three of them during collection (§6).

Method

Name That Dataset

The game in Figure 1 shows three images from each of twelve datasets: Caltech-101, UIUC, MSRC, Tiny Images, ImageNet, PASCAL VOC, LabelMe, SUN09, 15 Scenes, Corel, Caltech-256 and COIL-100 (footnote 1). To let a computer play, the authors sampled 1,000 images from the training portion of each dataset, trained a 12-way linear SVM, tested it on 300 images from each test set and repeated this 20 times. They compared four descriptors: $32 \times 32$ thumbnails in grayscale and in colour, gist, and a bag of HOG visual words (Figure 2 labels it Hog2x2). Figure 1’s images are the most discriminable of each dataset, those furthest from the SVM’s decision boundary; Figure 3 shows the reverse, the ImageNet and PASCAL images closest to the boundary of another dataset, which the authors call dataset look-alikes. To control for scene content, the same analysis was repeated on crops of cars from the five datasets that provided car bounding boxes: PASCAL, ImageNet, SUN09, LabelMe and Caltech-101.

Cross-dataset generalization

For the main experiment (§3.1) the authors chose six datasets that were in active use and shared annotated objects: SUN09, LabelMe, PASCAL VOC 2007, ImageNet, Caltech-101 and MSRC. All had bounding boxes, so two tasks were possible: classification, finding the images that contain the object, and detection, finding the boxes that contain it. The authors note that detection is classification with boxes as the images, but with far more, and more varied, negatives. They used the most standard methods of the time: Dalal and Triggs’s HOG detector with a linear SVM, and a bag-of-words classifier with a Gaussian-kernel SVM. The classes were “car” and “person”. Each classifier was trained on 500 positives and 2,000 negatives for classification, and on 100 positives and 1,000 negatives for detection, the largest numbers the smallest datasets allowed; tests used 50 positives and 1,000 negatives for classification and 10 positives and 20,000 negatives for detection, with results averaged over 20 runs. Performance is average precision. The authors warn that their protocol differs from each dataset’s own, so the absolute numbers mean little and the differences between them matter.

Negative-set bias

To ask whether a “not car” in PASCAL differs from a “not car” in MSRC (§3.2), each dataset’s detector was trained on its own positives and negatives, as before (100 and 1,000), and tested twice: once on its own test set, and once with its own positives but negatives drawn from all six datasets combined, keeping the number of negatives unchanged so that chance performance stayed the same. Tests used multiple runs of 10 positives against 20,000 negatives.

Sample value

Section 4 starts from learning curves. Figure 6 plots car detection AP against the number of training examples, from 3 to 200 on a log scale, for detectors trained on each of five datasets (MSRC is not included) and tested on SUN09, PASCAL, ImageNet and Caltech-101. The authors report no saturation and a linear relation between $\log n$ and $\log(1 - \mathrm{AP})$. Writing $\mathrm{AP}_i^j(n)$ for the AP of a detector trained on $n$ examples from dataset $i$ and tested on dataset $j$, they define the value $\alpha$ of a sample from $i$ on the benchmark of $j$ by

$$\mathrm{AP}_j^j(n) = \mathrm{AP}_i^j(n / \alpha),$$

so that one example from $i$ is worth $\alpha$ examples from $j$. On Figure 6 this is the horizontal shift that aligns the two curves, and the authors estimated it that way. Footnote 2 adds that an “incremental market value”, the help that other data gives after all the original data has been used, quickly converged to the same values.

Results

Datasets can be told apart

The best descriptor, the bag of HOG words, named the dataset 39% of the time against a chance rate of 1/12, or 8%, and its accuracy was still rising at 1,000 training images per dataset (Figure 2, left). In the authors’ labs, most people shown Figure 1 got more than 75% right. The confusion matrix (Figure 2, right) has a clear diagonal and groups scene-centric datasets apart from object-centric ones, with Caltech-101 and Caltech-256 confused with each other. The authors name two outliers: UIUC, whose test set differs from its training set, and COIL-100, a laboratory dataset. Read from the figure, COIL-100 is recognized 99% of the time, MSRC 60%, LabelMe 54% and 15 Scenes 51%, while UIUC’s test images are never assigned to UIUC. On the car crops the classifier still reached 61% against 20% chance. Figure 4 shows why: Caltech-101 cars are side views, ImageNet favours racing cars, PASCAL cars appear from non-canonical viewpoints, and SUN09 and LabelMe cars look alike, except that LabelMe cars are often occluded by small objects.

Cross-dataset generalization

Table 1 has a row for each training set and a column for each test set, followed by “Self” (the diagonal), “Mean others” and “Percent drop”, the relative fall from Self to Mean others. Its summary rows:

Task	Self (mean)	Mean others	Printed drop	Drop recomputed from these means
“car” classification	53.4	27.5	48%	48.5%
“car” detection	72.4	44.8	48%	38.1%
“person” classification	43.7	23.4	47%	46.5%
“person” detection	71.1	45.6	36%	35.9%

The car detection row prints the same 48% as the car classification row above it, but its own means give a 38% drop. Recomputing the per-dataset rows from the printed AP values reproduces every other printed drop to within one percentage point.

The authors’ reading (§3.1): the best results usually come from training and testing on the same dataset. Caltech-101 and MSRC are the easiest datasets to test on, by column averages, in every task. PASCAL and ImageNet generalize best most of the time, by row averages, although they score higher on the object-centric Caltech-101 and MSRC than on the scene-centric SUN09 and LabelMe. For car classification, a fall from 53.4 to 27.5 would, the authors write, turn a method ranked first in the PASCAL competition into one of the worst. Per training set, car classification drops range from 7% for PASCAL to 73% for Caltech-101, which scores 96.9 on its own test set and 26.7 on average elsewhere; car detection drops range from 4% (PASCAL) to 78% (Caltech-101). In person classification, PASCAL and ImageNet have negative drops (−31% and −63%) because their classifiers score higher on the easy Caltech-101 and MSRC test sets than on their own. Figure 5 shows a classifier trained on MSRC cars that finds cars only in MSRC. The authors call the results “rather depressing” and note that most of the datasets came from the same source, the Internet.

Four kinds of bias

To explain the drops, §3.1 names four likely causes:

• Selection bias: datasets prefer particular kinds of images, such as street scenes, nature scenes or keyword-search results, as the Name That Dataset experiment showed.
• Capture bias: photographers take pictures of objects in similar ways; the authors add that this bias might be similar across datasets.
• Category or label bias: semantic categories are often poorly defined, and labellers assign different labels to the same kind of object (“grass” or “lawn”, “painting” or “picture”, citing Malisiewicz and Efros, 2008).
• Negative-set bias: the negatives define what a dataset treats as the rest of the world, and an unrepresentative or unbalanced negative set can make classifiers overconfident and poorly discriminative. The authors judged this the least studied of the four.

The dataset bias article relates these to the statistical forms of dataset shift.

Negative sets

Detection		SUN09	LabelMe	PASCAL	ImageNet	Caltech-101	MSRC	Mean
“car”	own negatives	67.6	62.4	56.3	60.5	97.7	74.5	70.0
	all negatives	53.8	51.3	47.1	65.2	97.7	70.0	64.1
	drop	20%	18%	16%	−8%	0%	6%	8%
“person”	own negatives	67.4	68.6	53.8	60.4	100	76.7	71.1
	all negatives	52.2	58.0	42.6	63.4	100	71.5	64.6
	drop	22%	15%	21%	−5%	0%	7%	9%

Table 2 shows drops of 15–22% for SUN09, LabelMe and PASCAL, which the authors read as some of the other datasets’ negatives being confused with positives. The text says that ImageNet, Caltech-101 and MSRC show no drop, explaining that ImageNet’s negatives are already varied while Caltech-101 and MSRC are simply too easy; the table agrees for Caltech-101, shows ImageNet improving, and shows MSRC dropping by 6% for cars and 7% for persons. Whether a negative set is sufficient, rather than representative, the authors left open: a classifier for “boats” in a PASCAL-like task needs many images of water without boats to stop it from detecting the water, and testing sufficiency would need large amounts of unbiased labelled negatives, which they planned to gather with Mechanical Turk.

The market value of a sample

The paper’s example is that one LabelMe car is worth 0.26 PASCAL cars on the PASCAL benchmark. A modest gain, “maybe 10% AP”, for a detector trained on the 1,250 PASCAL VOC 2007 car samples would then need $1/0.26 \times 1250 \times 10 = 50{,}000$ LabelMe samples (§4; the product is about 48,000, and the paper does not explain the factor of 10). Table 3 gives the values among five datasets, each column a “market” in which the samples are priced:

One sample from	SUN09 market	LabelMe market	PASCAL market	ImageNet market	Caltech-101 market
SUN09	1	0.91	0.72	0.41	0
LabelMe	0.41	1	0.26	0.31	0
PASCAL	0.29	0.50	1	0.88	0
ImageNet	0.17	0.24	0.40	1	0
Caltech-101	0.18	0.23	0	0.28	1
Basket of currencies	0.41	0.58	0.48	0.58	0.20

Every off-diagonal value is below 1: a sample loses value when it is used on another dataset, which the authors attribute to the datasets’ biases, since nothing in theory requires it. The paper does not define the “basket of currencies” row; its values equal each column’s mean over all five rows, the market’s own samples included (our reading). In our reading of Figure 6, the zeros mark curves that never reach the target curve within the plotted range: no detector trained elsewhere approaches Caltech-101’s own AP on Caltech-101, and on PASCAL the curve of detectors trained on Caltech-101 cars levels off below 20 AP and then falls while PASCAL’s own curve keeps rising. The authors summarize the value of current datasets for training systems deployed in the real world as “better than nothing, but not by much”.

Verdict and recommendations

The discussion (§5) considers blaming the algorithms, since humans learn to see in a limited environment and still generalize, but argues that a dataset that defines “car” as the rear view of a race car leaves no reasonable algorithm able to call a side view of a family sedan a car. It judges Caltech-101 “extremely biased”, with virtually no observed generalization, and says it should have been retired long ago, as Ponce et al. (2006) had argued; MSRC also fared very poorly, while PASCAL VOC, ImageNet and SUN09 fared comparatively well.

The epilogue (§6) recommends running every new dataset through the paper’s battery of tests, for which the authors offered to publish their code and data, and suggests remedies for three kinds of bias:

• Selection bias. Automatically gathered datasets fared better than manually collected ones (the paper points to Figure 2), but keyword searches return particular kinds of images. Using several sources, such as several search engines in several countries as ImageNet did, helps somewhat; labelling a large unannotated collection by crowdsourcing might be better.
• Capture bias. Professional photographs and keyword-search results suffer most: objects sit in the centre, and an image search for “mug” returns handles facing right. Besides better sampling, the authors suggest flipping images left to right (noting that text then reads backwards), jittering them with small affine transformations, and automatic crops; data augmentation covers these transformations today.
• Negative-set bias. Add negatives from other datasets, as the paper did, or, following a suggestion of Mark Everingham, mine hard negatives from a large unlabelled set with standard algorithms such as bag of words and remove the true positives by hand, at the cost of a dataset biased against existing algorithms.

The authors close by admitting that, despite the title, their own biases have probably crept into the paper. The acknowledgements thank the Eyjafjallajökull volcano for the motivation to write it, and a disclaimer says the authors are listed in order of increasing procrastination ability.

Impact

The paper gave later work a vocabulary and a set of tests. Khosla et al. (2012) open by citing its experiments, with Ponce et al.’s, as evidence of selection, capture and negative-set bias, and Tommasi et al. (2015) write that it pointed out and named the main causes of dataset bias. Later papers reused its experiments. Khosla et al. (2012), with both authors among them, begin from its finding that adding data from other datasets can lower performance on a given test set, and reuse Name That Dataset to inspect their learned bias vectors. Zhou et al. (2014) cite it when testing scene classifiers across Places, SUN and ImageNet (§3.3), finding again that training and testing on the same dataset gave the best accuracy for a fixed number of training images. Tommasi et al. (2015) ran Name That Dataset on a 12-dataset testbed and the cross-dataset test on four of its collections, reporting its percentage drop alongside a measure of their own. Liu and He (2025) describe it as a call for a battle against dataset bias made just before the deep learning revolution, and revisit both of its main experiments.

The paper’s argument that a benchmark score describes a dataset as much as a method has a close relative in replication studies such as ImageNetV2, which rebuilt a test set by the original procedure to separate distribution shift from benchmark overfitting. Its verdict on Caltech-101 is recorded in the Caltech-101 article.

Limitations

The authors state several limits themselves: absolute AP values are not comparable with each dataset’s published results, only differences are (§3.1); their tests will not detect every source of bias (§6); the sufficiency of negative sets was not tested (§3.2); and their own biases have probably entered the paper. Others follow from the paper as printed:

• Small test sets. Detection AP was measured on 10 positives per run and classification on 50, averaged over 20 runs, and the tables report no spread, so small differences between cells cannot be judged; only the learning curves of Figures 2 and 6 carry error bars (our reading).
• Errors and disagreements. The car detection summary drop in Table 1 is printed as 48% where its own means give 38%. The text of §3.2 says that three datasets show no drop under pooled negatives, where Table 2 shows one unchanged, one improving and one dropping by 6–7%. The 2010 significance result is attributed to an IJCV paper that analyses VOC2007.
• A relative measure. The percentage drop hides the level it falls from. Tommasi et al. (2015) point out that a 75% drop from 100 AP differs from a 75% drop from 25 AP, and propose a measure based on the absolute difference between Self and Mean others.
• Extrapolated prices. Market values are horizontal shifts between learning curves measured up to 200 training examples (Figure 6), while the 50,000-sample example starts from 1,250 PASCAL samples, beyond the plotted range, and relies on an unexplained factor of 10 (our reading).
• One generation of features. Every experiment used hand-designed features (thumbnails, gist, bags of HOG words, HOG templates) with SVMs; whether learned features remove the effect was a question for later work.
• Two classes. The cross-dataset and negative-set experiments use only “car” and “person”, the classes shared by all six datasets.

What Came After

Undoing the damage. Khosla et al. (2012) treated each dataset as a biased view of one visual world and learned, in a max-margin model, visual-world weights shared by all datasets plus a bias vector for each, so that the classifier for dataset $i$ is $w_i = w_{vw} + \Delta_i$. On a dataset held out from training, the visual-world weights beat an SVM trained on the pooled data by 2.8 points of mean AP in classification and 0.7 in detection (their Figs. 2 and 3).

Learned features. Tommasi et al. (2015, arXiv version) repeated Name That Dataset and the cross-dataset test with DeCAF features from a convolutional network. The datasets became easier to tell apart than with bag-of-SIFT features, cross-dataset gaps narrowed for some classes and datasets and widened for others, and negative-set bias persisted whatever the features. Neither Khosla et al.’s method nor the domain adaptation methods they tested helped on top of DeCAF, while self-labelling the target data did.

A decade later. Liu and He (2025) played Name That Dataset with six large-scale datasets of the pre-training era (YFCC, CC, DataComp, WIT, LAION and ImageNet), sampling 1 million training and 10,000 validation images from each. A ConvNeXt-T network told YFCC, CC and DataComp apart with 84.7% accuracy against 33.3% chance; across all 20 combinations of three of their six datasets accuracy was above 62%, and above 80% in 16 of them (Table 2). AlexNet reached 77.8% on the same three datasets (Table 3), while 20 human volunteers with machine-learning backgrounds averaged 45.4%. Their cross-dataset test, using self-supervised validation loss in place of AP because the datasets share no labelled task, again found each dataset best served by its own training data, with a training set mixed from all six best on average (Table 9). For bias in datasets of people, replication studies and later mitigation work, see dataset bias.