Explaining Automatic Image Assessment

Max Lisaius; Scott Wehrwein

REVIEW 3 major objections 5 minor 28 references

Explaining Automatic Image Assessment

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper's claim is that aesthetic quality can be explained by training the same scoring network on separate visual channels—color, depth, saliency, and blur—and comparing which channel best predicts human scores.

desk verdict The dataset-mean threshold critique is real and citable; the modality-explainability story is overstated and internally contradicted, but the paper deserves a serious referee for the baseline finding alone. read the letter →

arxiv 2502.01873 v1 pith:RYKL33JS submitted 2025-02-03 cs.CV

Max Lisaius , Scott Wehrwein This is my paper · ORCID

classification cs.CV

keywords imageaestheticsexplainableAIsaliencymapsdepthestimationearthmover'sdistancetransferlearningcovarianceanalysisAVAdataset

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to make automatic image-aesthetic scores explainable without hand-labeled categories. The authors take a standard aesthetic-scoring network, retrain it on four versions of the same large photo set (original color, estimated depth, estimated saliency, and a blurred copy), and compare how well each version predicts human ratings. Saliency comes out clearly ahead of depth and blur, close behind full color, which suggests that where viewers look carries more aesthetic weight than geometric layout or fine detail. Covariance analysis against existing category labels lets the method single out trends, such as all models struggling on human subjects and depth aligning with animal or foreground-shape content. The payoff is a fully automatic route to aesthetic explanation: instead of labeling what makes a photo good, train on what remains when other information is stripped away.

What carries the argument

The load-bearing mechanism is the modality decomposition: each image in the AVA dataset is converted into an estimated depth map (single-image depth estimation), an estimated saliency map (OpenCV saliency), and a heavily blurred/resized version, and the same NIMA-style network is trained on each channel separately, with partial layer freezing used to transfer from RGB pretraining. The comparison is carried by earth mover's distance as the training objective, modified by multiplicative mean- and variance-weighting terms, plus evaluation metrics (two-class accuracy at two thresholds, LCC, SRCC, EMD, and histogram-EMD between predicted and ground-truth mean/variance distributions). The covariance analysis then relates each modality's per-image EMD to existing category labels, turning a scalar score into a vector of category-level affinities.

What would settle it

Rerun the modality comparison on images with eye-tracking-derived saliency maps and depth from RGB-D sensors or LiDAR; if saliency's advantage over depth and blur shrinks or reverses relative to the estimated-map version, the reported ranking reflects estimator quality rather than aesthetic content.

Watch

Extended reading notes

Core claim

The central discovery is a measurement, not just a model: across two-class accuracy, rank correlation, EMD, and the shape of predicted score distributions, a network trained on saliency maps of the AVA dataset outperforms networks trained on depth maps or blurred images, reaching 75.77% two-class accuracy at the standard threshold and 70.11% at the dataset-mean threshold. The paper treats this as evidence that visual saliency is the most informative non-color channel for aesthetics, and that the modality comparison itself is an explanation tool: images where one modality predicts best form coherent visual groups, and covariance with external labels (EVA aesthetics/attributes, AVA challenge categories, CADB composition categories) recovers interpretable relationships. A secondary discovery is that the authors' re-weighted EMD loss—standard EMD multiplied by terms that grow with the distance of the predicted mean and variance from dataset constants—improves linear and rank correlation over the original NIMA model on RGB images, while trading off raw EMD.

Load-bearing premise

The depth and saliency inputs are produced by automatic estimators, not measured ground truth, and the paper's conclusions about which visual channel matters most assume those estimates faithfully represent depth and visual interest rather than artifacts of the estimators.

Editorial extensions

If this is right

A saliency-only network can serve as an approximate explainability probe: if saliency alone recovers most of the aesthetic signal, then aesthetic models are largely responding to where people look, not to color or fine texture.
The automatic grouping of images by best-predicted modality produces visual aesthetic categories without any manual labeling, and these groups differ in interpretable ways (e.g., depth preferring strong foreground shapes, blur being composition-agnostic).
The modified EMD loss offers a tunable tradeoff: with the mean and variance terms active, correlation with human ratings improves over the base NIMA model, but pure EMD loss remains best for distributional distance—so future work can tune for the metric that matters.
Threshold choice in two-class accuracy matters enough that the paper recommends reporting accuracy at the dataset mean (5.38) in addition to the conventional 5.0; otherwise a trivial all-good baseline scores above 70%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the authors' modality logic is right, adding a pure grayscale (luminance-only) modality would isolate whether saliency's advantage comes from attentional structure or simply from dropping color; that experiment is not reported but follows directly.
The result is conditional on the off-the-shelf depth and saliency estimators; replacing them with stronger modern estimators could change the modality ranking, so the paper's saliency-first conclusion should be read as an upper bound for these particular estimates.
The consistently weak performance on human subjects across modalities suggests a bias in the AVA scoring process itself, which could be tested by re-weighting the validation set for subject content and checking whether the negative covariances persist.
The covariance analysis could be turned into an auditing tool: given any proposed aesthetic attribute, the same per-category EMD comparison would show whether a model's errors align with that attribute without needing new labels.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

The thing worth knowing about this paper: it has one genuinely useful result buried in a lot of exploratory noise. The baseline critique is correct—AVA's mean is 5.3833, so a model that predicts 'good' for every image gets 70.82% at the standard threshold of 5. That's a real finding that should change how people report two-class accuracy on this benchmark. The rest of the paper—training separate NIMA-style models on depth, saliency, and blur versions of AVA and comparing their metrics and covariances with labeled categories—is an interesting idea, and the authors deserve credit for a clear experimental setup and for including covariance analyses that go beyond simple accuracy.

The soft spots are substantial. The conclusion credits the modified loss terms for beating NIMA on RGB correlation, but the paper's own controlled sweep (Table 5) shows pure EMD gives the best LCC/SRCC and each added term makes it worse. That's a direct internal contradiction. Similarly, the discussion says the saliency modality outperforms all others on all metrics, but Tables 1–3 show RGB is better on every reported metric. If 'modalities' means the three generated versions, excluding RGB, that needs to be explicit; as written it looks like an error. The stress-test concern about estimated depth and saliency maps is also on point: those maps come from off-the-shelf estimators, are not validated on AVA, and the authors concede they fail on post-processed or bordered images. If estimator error correlates with aesthetics, the saliency>blur>depth ranking could reflect the estimator rather than the perceptual feature. That's a threat to the central explainability claim that isn't addressed. Finally, there's no code, no final hyperparameters, no splits, seeds, or error bars, which limits reproducibility.

None of this kills the threshold critique, which stands independently. And the modality comparison, while inconclusive, is a reasonable dataset that a revised paper could present more carefully. The overclaims are fixable with rephrasing and a clearer separation of 'generated modalities' from the RGB baseline, but the estimator-fidelity problem needs more than wording—it needs a control or robustness analysis before the explainability story is credible.

I'd send this to peer review rather than desk-reject, because the baseline insight deserves a venue and a good referee can sort the wheat from the chaff. I wouldn't publish it as is. I'd cite the threshold result, and it's a decent reading-group topic for a discussion of evaluation practices.

Best, [Your name]

Referee Report

3 major / 5 minor

Summary. The paper proposes an explainability method for automatic image aesthetic assessment. The authors reimplement the NIMA architecture with a modified EMD loss that includes mean- and variance-deviation penalty terms whose constants (5.3833 and 2.084) are derived from the AVA dataset statistics. They train models on the original RGB images and on three derived modalities (estimated depth, computed saliency, and blurred versions), and compare them with 2-class accuracy (using both the conventional 5.0 threshold and the dataset-mean threshold), LCC, SRCC, and EMD. They report that saliency is the best non-RGB modality, that RGB achieves the best overall scores, and that the additional loss terms improve correlation over NIMA. They also present covariance analyses relating modality-specific EMD performance to category labels in EVA, AVA, and CADB. The paper concludes that this automatic modality comparison can explain aesthetic assessment by revealing which visual information drives aesthetic scores.

Significance. The paper makes one clearly sound and useful observation: the conventional 5.0 threshold in AVA is unwarranted because the dataset mean is 5.3833, so a trivial always-good baseline already achieves 70.82% accuracy; reporting a mean-thresholded accuracy is a fair and easily verifiable point. The general idea of comparing networks trained on different visual modalities as a form of automatic explanation is also interesting and could be valuable as a diagnostic tool. However, the two central claims that would make the paper significant are not supported. First, the loss-term improvement claim is contradicted by the paper's own ablation table. Second, the modality ranking is built on estimated depth and saliency maps that are never validated, so the observed ranking could be an artifact of the estimators rather than a property of the aesthetic features. As a result, the claimed explainability contribution is not established. The covariance analyses are exploratory but lack statistical significance testing, further limiting their evidentiary value.

major comments (3)

[Section 4.4 (Table 5) and Section 7 (Conclusion)] The Conclusion states: "We show improvements in correlation scores over NIMA using hyperparameter controlled loss terms," and Section 4.4 claims the loss terms improved metrics. However, Table 5, which is the controlled comparison of Pure EMD versus EMD*∆µ, EMD*∆σ2, and EMD*∆µ*∆σ2, shows that Pure EMD achieves the highest LCC and SRCC (0.4994/0.4918) while the combined loss achieves the lowest (0.4309/0.4244). The added loss terms monotonically degrade correlation. The paper's own data therefore contradicts the central claim about the loss terms. The final RGB model in Table 2 may have been trained with loss terms, but the ablation in Table 5 cannot support the claim that the loss terms are responsible for any improvement; if anything, it indicates the opposite.
[Section 3.3 and Section 5 (Modalities and Discussion)] The central claim that comparing modality-specific models reveals which visual information drives aesthetic scores is ungrounded because the depth and saliency modalities are produced by off-the-shelf estimators (OpenCV saliency and a single-image depth estimator) and are never validated on the AVA domain. The paper concedes in Section 5 that the depth modality struggles with post-processed images and that multiple modalities fail on bordered or watermarked images. Since AVA is a large, noisy web-collected corpus, such cases are likely present. If estimator error is correlated with aesthetic score—e.g., saliency maps highlighting high-contrast regions that are also judged pleasing—then the observed ranking (saliency > blur > depth) reflects the estimators, not the intended aesthetic properties. No control experiment, ablation, or fidelity analysis separates these possibilities, so the primary explanatory conclusion is not supported.
[Section 3.2 and Section 4.3 (Loss constants and hyperparameter tuning)] The loss constants 5.3833 and 2.084 are stated to be the average mean and variance of the ground-truth distributions in the AVA dataset. The paper does not specify whether these constants are computed from the training split only or from the full dataset, including the test set. Additionally, all hyperparameter sweeps (loss weights, layer-freeze fractions, learning rates) are performed on the same AVA benchmark, and only a partial external AADB check is reported. This creates a risk that the reported gains over NIMA on AVA are in part the result of fitting dataset-specific statistics and hyperparameters to the evaluation benchmark. The authors should clarify the split used for computing the constants and report results with hyperparameters chosen without access to AVA test statistics.

minor comments (5)

[Section 3.3] The blur modality pipeline is described only in words; providing a pseudocode block or a precise parameter summary with a figure would improve reproducibility.
[Section 4.3] "with SDG" should be "with SGD" (stochastic gradient descent); also, the statement that hyperparameters are tuned with Bayesian optimization is not accompanied by the search ranges or the final selected values, which makes reproducing the training setup difficult.
[Section 4.4 (Tables 1–3)] The differences between modality scores (e.g., saliency vs. blur) are small and no statistical significance tests, confidence intervals, or multiple-run variability are reported; these should be added to support ranking claims.
[Section 4.5] The covariance values in Figures 11, 12, 14, and 15 are not reported numerically, and the text describes only qualitative trends; reporting the values with a test for significance (e.g., permutation test) would be more informative.
[Throughout] There are several typos and formatting inconsistencies, including "aestheic" in Section 1 and inconsistent reference formatting for [1] and [11]; a careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's comparisons are empirical and self-contained; the loss constants are training hyperparameters, not predictions.

full rationale

The paper's central derivation is an empirical pipeline: train NIMA-style regressors on RGB, depth, saliency, and blur versions of AVA, then compare EMD, LCC, SRCC, and two-class accuracy. The inputs (AVA ground-truth distributions, OpenCV saliency maps, MINE depth estimates, blurred images) are distinct from the outputs (per-modality accuracies, correlation scores, and covariance trends). No quoted step exhibits an equation that reduces to its own input by construction. The modified loss uses dataset-wide constants (5.3833 and 2.084) taken from AVA ground-truth statistics, but these are fixed weights in the loss function, not predicted values; moreover, the authors explicitly state 'For all validation and metric purposes, the original EMD loss function was preserved so that hyperparameters have no effect on reported EMD metrics,' so the reported EMD scores are not forced by the loss constants. Hyperparameter sweeps on AVA are standard model selection rather than a fitted parameter being reported as a prediction, and the paper includes an out-of-distribution check on AADB. The acknowledged limitations around estimated depth and saliency maps (Section 5) are external-validity threats, not circularity: the modality comparison could be confounded by estimator artifacts, but that is a correctness risk, not a definitional reduction. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed known results. The paper is self-contained against external benchmarks, so the appropriate circularity score is 0.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The paper rests on the AVA label distribution, on off-the-shelf depth and saliency estimators, and on the assumption that per-modality network comparisons expose aesthetic content. The loss is anchored by two AVA-derived constants and several unreported hyperparameters. No novel entities are postulated.

free parameters (7)

h_mu (loss weight for |mean deviation|)
Tunable multiplier in Loss = EMD * h_mu * h_sigma2; searched by Bayesian optimization, final value not reported.
h_sigma2 (loss weight for |variance deviation|)
Second tunable multiplier in the combined loss; final value not reported.
AVA dataset mean constant 5.3833 = 5.3833
Computed from the ground-truth distributions of the AVA data used for training and evaluation; inserted into the loss as a fixed constant.
AVA dataset variance constant 2.084 = 2.084
Computed from the same AVA ground-truth distributions; used in the Delta_sigma2 loss term.
Layer freeze fraction for transfer learning = 85-95 percent early, final not reported
Hyperparameter controlling how many late layers keep RGB pretrained weights; tuned per modality, exact final values not given.
Blur modality pipeline parameters = 9x9 and 3x3 Gaussian, bilateral d=20 sigma=50, 32x32 downscale
Hand-chosen smoothing recipe; no ablation shows how these exact parameter choices affect the modality ranking.
Learning rate and decay schedule
Bayesian sweeps tuned learning rates and decay for SGD and Adam fine-tuning, but final values are not reported.

assumptions (6)

domain assumption AVA ground-truth score distributions are a valid measure of aesthetic quality.
All training and evaluation assume the AVA labels, collected over a decade ago, are the ground truth for aesthetics; the paper itself later questions dataset-age bias.
domain assumption The estimated depth maps and saliency maps faithfully represent geometric composition and visual interest.
Depth comes from a single-image estimator [16] and saliency from OpenCV; the authors admit these are estimates and fail on post-processed or watermarked images (Section 5).
domain assumption The NIMA architecture with VGG backbone and EMD loss is an adequate base model for probing aesthetic features.
The paper selects NIMA to avoid attention mechanisms, implicitly assuming its learned features are representative of aesthetic assessment (Section 3.2).
ad hoc to paper Training a separate network per modality and comparing final metrics reveals which visual information drives aesthetic scores.
This is the paper's central linking hypothesis; if frozen layers or shared pretraining force the networks to converge to RGB features, modality differences would be masked. The authors note modalities picked up common content (Section 4.4).
standard math The squared CDF-based EMD with r=2 is a suitable differentiable distance for score distributions.
Adopted from [1] and [11]; needed for loss and metric comparisons.
domain assumption Transfer learning by freezing late layers preserves useful RGB-learned aesthetic features while allowing early layers to adapt.
The freezing scheme is introduced to escape local minima; if it prevents modality-specific learning, the modality comparisons are biased (Section 3.4).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Explaining Automatic Image Assessment." pith.science (2026). https://pith.science/paper/RYKL33JS

@misc{pith2026250201873,
  author       = {Pith},
  title        = {Pith review of: Explaining Automatic Image Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RYKL33JS}},
  note         = {Machine review of arXiv:2502.01873}
}

read the original abstract

Previous work in aesthetic categorization and explainability utilizes manual labeling and classification to explain aesthetic scores. These methods require a complex labeling process and are limited in size. Our proposed approach attempts to explain aesthetic assessment models through visualizing dataset trends and automatic categorization of visual aesthetic features through training neural networks on different versions of the same dataset. By evaluating the models adapted to each specific modality using existing and novel metrics, we can capture and visualize aesthetic features and trends.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

[1]

modalities

INTRODUCTION Assessment of image aesthetics is a subjective and difficult problem in computer vision. Recently, some success has been demonstrated using neural networks to evaluate the aesthetic quality of an image; however, these approaches usually lack the ability to explain what makes an image aesthetically good or bad. More explainable aesthetic asses...

work page
[2]

Older methods that use saliency and rule of thirds based techniques [4] used limited datasets and were implemented using algorithms

RELATED WORK Image aesthetic and composition evaluation is an open re- search subject with ongoing applications and a history of in- terest in the field. Older methods that use saliency and rule of thirds based techniques [4] used limited datasets and were implemented using algorithms. Other methods in related ar- eas aim to recompose or improve images [5...

work page
[3]

modalities

METHODS 3.1. NIMA In 2018 NIMA [1] was published and introduced a new loss function to the task, earth movers distance, a loss function originally proposed in [11]. Given that the dataset being used contained a distribution of scores it made sense to have the network also predict a distribution, therefore EMD was a log- ical choice to measure the distance...

work page 2018
[4]

Datasets The dataset we chose to apply our methods to is the most widely used image aesthetics dataset; A V A or Aesthetic Vi- sual Analysis [6]

EXPERIMENTS 4.1. Datasets The dataset we chose to apply our methods to is the most widely used image aesthetics dataset; A V A or Aesthetic Vi- sual Analysis [6]. This dataset contains over 255,000 images with many aesthetic assessment scores per image making it the largest dataset available for this task. These assessment scores are compiled into a 10-bi...

work page
[5]

It was surpris- ing to find such significant similarities between the different modalities in preference

DISCUSSION From our experiments and results we are able to make some conclusions about automatic aesthetic assessment and the modality breakdown methods we applied. It was surpris- ing to find such significant similarities between the different modalities in preference. As we observed in both the EMD by score and the covariance analysis, the differences s...

work page
[6]

This would allow us to explore in greater depth the differences in modalities if they are present

FUTURE WORK In the future we hope to apply our method to larger and more recent models supported by a corresponding increase in ac- cess to hardware. This would allow us to explore in greater depth the differences in modalities if they are present. If these models also fail to show a large macro difference be- tween modalities, it would solidify our exper...

work page
[7]

Baseline

LCC (linear correlation coefficient) and SRCC (Spearman rank correlation coefficient) are useful metrics for determin- ing the correlation of a model’s performance to the ground truth. Earth movers distance is also an available metric for the task, and given it is the basis for the loss function we used for training, we also evaluate the models using pure...

work page
[8]

Attention- based multi-patch aggregation for image aesthetic as- sessment,

Kekai Sheng, Weiming Dong, Chongyang Ma, Xing Mei, Feiyue Huang, and Bao-Gang Hu, “Attention- based multi-patch aggregation for image aesthetic as- sessment,” in Proceedings of the 26th ACM Interna- tional Conference on Multimedia, New York, NY , USA, 2018, MM ’18, p. 879–886, Association for Computing Machinery

work page 2018

Show all 28 references

[9]

Pool-3FC [10] pools features in a arXiv:2502.01873v1 [cs.CV] 3 Feb 2025 comparable way to also squeeze the most amount of informa- tion out of the source data

make use of patches, attention mechanisms, and signif- icant hardware to focus in on specific aesthetic features at- tacking accuracy metrics. Pool-3FC [10] pools features in a arXiv:2502.01873v1 [cs.CV] 3 Feb 2025 comparable way to also squeeze the most amount of informa- tio...

2025 arXiv
[10]

We show im- provements in correlation scores over NIMA using hyper- parameter controlled loss terms, and apply this method to generated modalities of the source dataset

CONCLUSION Our method provides a unique and novel insight into a dom- inant and widely accepted aesthetics dataset. We show im- provements in correlation scores over NIMA using hyper- parameter controlled loss terms, and apply this method to generated modalities of the source ...
[11]

NIMA: neural image assessment,

Hossein Talebi Esfandarani and Peyman Milanfar, “NIMA: neural image assessment,” CoRR, vol. abs/1709.05424, 2017

2017 arXiv
[12]

Image composi- tion assessment with saliency-augmented multi-pattern pooling,

Bo Zhang, Li Niu, and Liqing Zhang, “Image composi- tion assessment with saliency-augmented multi-pattern pooling,” arXiv preprint arXiv:2104.03133, 2021

2021 arXiv
[13]

Eva: An explainable visual aesthetics dataset,

Chen Kang, Giuseppe Valenzise, and Fr ´ed´eric Dufaux, “Eva: An explainable visual aesthetics dataset,” Joint Workshop on Aesthetic and Technical Quality Assess- ment of Multimedia and Media Analytics for Societal Trends, 2020

2020
[14]

Automatic aesthetic photo composition,

Roberto Gallea, Edoardo Ardizzone, and Roberto Pir- rone, “Automatic aesthetic photo composition,” 09 2013, vol. 8157, pp. 21–30

2013
[15]

A framework for photo-quality assess- ment and enhancement based on visual aesthetics,

Subhabrata Bhattacharya, Rahul Sukthankar, and Mubarak Shah, “A framework for photo-quality assess- ment and enhancement based on visual aesthetics,” in Proceedings of the 18th ACM International Conference on Multimedia, New York, NY , USA, 2010, MM ’10, p. 271–280, Associatio...

2010
[16]

Ava: A large-scale database for aesthetic visual analy- sis,

Naila Murray, Luca Marchesotti, and Florent Perronnin, “Ava: A large-scale database for aesthetic visual analy- sis,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 2408–2415

2012
[17]

good” or “bad

datasets. The EV A dataset is a subset of the A V A dataset that has been relabeled along with additional category labels to attempt to explain the aesthetic content. These categories include ”color and light”, ”composition and depth”, ”qual- ity”, and ”semantic” for aesthetic...
[18]

Deep multi-patch aggregation network for image style, aesthetics, and quality estimation,

Xin Lu, Zhe Lin, Xiaohui Shen, Radomir Mech, and James Z. Wang, “Deep multi-patch aggregation network for image style, aesthetics, and quality estimation,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015

2015
[19]

A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment,

Shuang Ma, Jing Liu, and Chang Wen Chen, “A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment,” CoRR, vol. abs/1704.00248, 2017

2017 arXiv
[20]

Ef- fective aesthetics prediction with multi-level spatially pooled features,

Vlad Hosu, Bastian Goldl ¨ucke, and Dietmar Saupe, “Ef- fective aesthetics prediction with multi-level spatially pooled features,” CoRR, vol. abs/1904.01382, 2019

1904 arXiv
[21]

Squared earth mover’s distance-based loss for training deep neu- ral networks,

Le Hou, Chen-Ping Yu, and Dimitris Samaras, “Squared earth mover’s distance-based loss for training deep neu- ral networks,” CoRR, vol. abs/1611.05916, 2016

2016 arXiv
[22]

Rethinking the inception architecture for computer vision,

Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna, “Rethinking the inception architecture for computer vision,” CoRR, vol. abs/1512.00567, 2015

2015 arXiv
[23]

Mobilenets: Efficient convolutional neural networks for mobile vision appli- cations,

Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision appli- cations,” CoRR, vol. abs/1704.04861, 2017

2017 arXiv
[24]

Very deep convolutional networks for large-scale image recogni- tion,

Karen Simonyan and Andrew Zisserman, “Very deep convolutional networks for large-scale image recogni- tion,” CoRR, vol. abs/1409.1556, 2015

2015 arXiv
[25]

Mean-variance loss for deep age estimation from a face,

Hongyu Pan, Hu Han, Shiguang Shan, and Xilin Chen, “Mean-variance loss for deep age estimation from a face,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

2018
[26]

Single-view view synthesis with multiplane images,

Richard Tucker and Noah Snavely, “Single-view view synthesis with multiplane images,” in The IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), June 2020

2020
[27]

Photo aesthetics ranking net- work with attributes and content adaptation,

Shu Kong, Xiaohui Shen, Zhe Lin, Radom ´ır Mech, and Charless C. Fowlkes, “Photo aesthetics ranking net- work with attributes and content adaptation,”CoRR, vol. abs/1606.01621, 2016

2016 arXiv
[28]

Adam: A method for stochastic optimization,

Diederik P. Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” CoRR, vol. abs/1412.6980, 2015

2015 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

modalities

INTRODUCTION Assessment of image aesthetics is a subjective and difficult problem in computer vision. Recently, some success has been demonstrated using neural networks to evaluate the aesthetic quality of an image; however, these approaches usually lack the ability to explain what makes an image aesthetically good or bad. More explainable aesthetic asses...

work page

[2] [2]

Older methods that use saliency and rule of thirds based techniques [4] used limited datasets and were implemented using algorithms

RELATED WORK Image aesthetic and composition evaluation is an open re- search subject with ongoing applications and a history of in- terest in the field. Older methods that use saliency and rule of thirds based techniques [4] used limited datasets and were implemented using algorithms. Other methods in related ar- eas aim to recompose or improve images [5...

work page

[3] [3]

modalities

METHODS 3.1. NIMA In 2018 NIMA [1] was published and introduced a new loss function to the task, earth movers distance, a loss function originally proposed in [11]. Given that the dataset being used contained a distribution of scores it made sense to have the network also predict a distribution, therefore EMD was a log- ical choice to measure the distance...

work page 2018

[4] [4]

Datasets The dataset we chose to apply our methods to is the most widely used image aesthetics dataset; A V A or Aesthetic Vi- sual Analysis [6]

EXPERIMENTS 4.1. Datasets The dataset we chose to apply our methods to is the most widely used image aesthetics dataset; A V A or Aesthetic Vi- sual Analysis [6]. This dataset contains over 255,000 images with many aesthetic assessment scores per image making it the largest dataset available for this task. These assessment scores are compiled into a 10-bi...

work page

[5] [5]

It was surpris- ing to find such significant similarities between the different modalities in preference

DISCUSSION From our experiments and results we are able to make some conclusions about automatic aesthetic assessment and the modality breakdown methods we applied. It was surpris- ing to find such significant similarities between the different modalities in preference. As we observed in both the EMD by score and the covariance analysis, the differences s...

work page

[6] [6]

This would allow us to explore in greater depth the differences in modalities if they are present

FUTURE WORK In the future we hope to apply our method to larger and more recent models supported by a corresponding increase in ac- cess to hardware. This would allow us to explore in greater depth the differences in modalities if they are present. If these models also fail to show a large macro difference be- tween modalities, it would solidify our exper...

work page

[7] [7]

Baseline

LCC (linear correlation coefficient) and SRCC (Spearman rank correlation coefficient) are useful metrics for determin- ing the correlation of a model’s performance to the ground truth. Earth movers distance is also an available metric for the task, and given it is the basis for the loss function we used for training, we also evaluate the models using pure...

work page

[8] [8]

Attention- based multi-patch aggregation for image aesthetic as- sessment,

Kekai Sheng, Weiming Dong, Chongyang Ma, Xing Mei, Feiyue Huang, and Bao-Gang Hu, “Attention- based multi-patch aggregation for image aesthetic as- sessment,” in Proceedings of the 26th ACM Interna- tional Conference on Multimedia, New York, NY , USA, 2018, MM ’18, p. 879–886, Association for Computing Machinery

work page 2018