Towards Unsupervised Image Harmonisation

Alan Dolhasz

, Carlo Harvey

and Ian Williams

Digital Media Technology Lab, Birmingham City University, Birmingham, U.K.

Keywords:

Image Compositing, Harmonisation, Artifact Detection, End-to-End Compositing, Deep Learning.

Abstract:

The ﬁeld of image synthesis intrinsically relies on the process of image compositing. This process can be auto-

matic or manual, and depends upon artistic intent. Compositing can introduce errors, due to human-detectable

differences in the general pixel level transforms of component elements of an image composite. We report on

a pilot study evaluating a proof-of-concept automatic image composite harmonisation system consisting of a

state-of-the-art deep harmonisation model and a perceptually-based composite luminance artifact detector. We

evaluate the performance of both systems on a large data-set of 68128 automatically generated image com-

posites and ﬁnd that without any task-speciﬁc adaptations, the end-to-end system achieves comparable results

to the baseline harmoniser fed with ground truth composite masks. We discuss these ﬁndings in the context of

extending this to an end-to-end, multi-task system.

1 INTRODUCTION

Image compositing is a common task in image pro-

cessing where an object from one image is extracted

and inserted into another image, referred to as the

scene, with the aim of creating a plausible, realis-

tic result (Wright, 2013). Due to inherent dispar-

ities in appearance between the object and scene,

commonly resulting from differences in illumination,

camera intrinsics, post-processing, encoding or com-

pression, component elements of a composite often

require post-processing in order to create a com-

pelling and realistic ﬁnal result. To address these

issues, a wide range of automatic compositing tech-

niques have been proposed. These include alpha mat-

ting - linear combinations of object and scene pixel

values (Porter and Duff, 1984), gradient-domain op-

timization techniques (P

erez et al., 2003; Agarwala

et al., 2004; Levin et al., 2004), visual appearance

transfer (Reinhard et al., 2001; Lalonde and Efros,

2007) and multi-scale methods (Burt and Adelson,

1983a,b; Sunkavalli et al., 2010).

More recently, deep learning (DL) based ap-

proaches have achieved considerable success in the

domain of image compositing. Notably Tsai et al.

(2017) adopt the denoising autoencoder (DAE) (Vin-

cent et al., 2008) attempting to learn the composit-

https://orcid.org/0000-0002-6520-8094

https://orcid.org/0000-0002-4809-1592

https://orcid.org/0000-0002-0651-0963

ing function directly from image data, including se-

mantic information derived from ground truth seman-

tic segmentation labels. Chen and Kae (2019) lever-

age a generative adversarial network (GAN) to learn

both colour-based and geometric transformations in

order to perform compositing of arbitrary objects into

arbitrary scenes. Conditional GANs have also been

adopted to address this problem, by learning to model

joint distributions of different object classes and their

interactions in image space (Azadi et al., 2018), as

well as performing colour and gradient blending be-

tween composite elements, which have been semanti-

cally aligned.

These existing methods are not without limita-

tions. Firstly, they focus on creation of new compos-

ites, thus requiring object/scene segmentation masks

to be available at input. This limits their use for

cases where these are not available, such as improve-

ment of existing image composites. Secondly, they

do not explicitly leverage human perceptual charac-

teristics, such as their sensitivity to various image

artifacts or magnitude of mismatch between object

and scene (Dolhasz et al., 2016). Finally, the masks

supplied to such algorithms provide only a binary

indication whether a given pixel belongs to the ob-

ject or the scene. This implies the entire region re-

quires correction and induces a generic transforma-

tion, such as colour transfer, uniformly across the re-

gion. This can result in the harmonisation algorithm

over-compensating and generating a suboptimal out-

574

Dolhasz, A., Harvey, C. and Williams, I.

Towards Unsupervised Image Harmonisation.

DOI: 10.5220/0009354705740581

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages

574-581

ISBN: 978-989-758-402-2; ISSN: 2184-4321

put, even compared to the unprocessed input compos-

ite.

We argue that perceptual detection of compos-

ite artifacts should be explicitly modelled in deep-

learning-based image compositing and harmonisa-

tion. Our reasoning behind this is as follows. Firstly,

it enables design of end-to-end harmonisation sys-

tems without the need for manually supplied ob-

ject masks, allowing harmonisation of composites for

which masks are not available. Secondly, the explicit

encoding of the location and perceptual magnitude

of errors in the output of the model allows the pro-

cess to take advantage of the beneﬁts of multi-task

learning in terms of generalisation (Ruder, 2017; Ran-

jan et al., 2017). The potential applications of such

automatic compositing systems are wide-ranging, in-

cluding improvement of legacy content, detection of

image manipulations and forgery, perceptually-based

metrics and image synthesis.

Consequently, in this study, we design a proof-

of-concept end-to-end compositing pipeline consist-

ing of a detector network, which outputs masks cor-

responding to regions requiring harmonisation, and a

harmoniser network, which corrects the detected re-

gions. We then assess the impact of object masks

predicted by the detector on the accuracy of the

harmoniser, compared to using ground truth object

masks. Our study adopts two existing networks

- the Deep Harmonisation algorithm proposed by

Tsai et al. (2017) as the harmoniser network, and a

perceptually-based fully convolutional network pro-

posed by Dolhasz et al. (2019) as the detector net-

work.

We show that our prototype end-to-end system,

using the the detector network without any task-

speciﬁc adaptations or re-training, produces results

which are comparable to those obtained using ground

truth masks. To our knowledge this is the ﬁrst work

investigating the combination of a deep-learning-

based detection model with a composite harmonisa-

tion one to both detect and ﬁx composites. We are

currently developing a complete, end-to-end version

of the model, trained speciﬁcally for this purpose.

2 RELATED WORK

2.1 Image Compositing &

Harmonisation

Automatic image compositing and harmonisation are

both active and challenging problems in the domain

of image understanding and processing. Image com-

positing concerns the entire process of combining

regions from different source images into a plausi-

ble whole, while image harmonisation focuses on the

problem of matching the various appearance features

between the object and scene, such as noise, contrast,

texture or blur, while assuming correctly aligned ge-

ometric and illumination properties (Sunkavalli et al.,

2010).

Similarly to the problem of image in-painting,

compositing and harmonisation are both ill-posed

problems (Guillemot and Le Meur, 2013). For a given

region requiring correction many different arrange-

ments of pixels could be deemed plausible. This is

in contrast to problems where the solution is unique.

Depending on the content and context of an image

composite, some scene properties, and thus required

object corrections, may be inferable from the infor-

mation contained within the image or its metadata,

such as the characteristics of the illuminant (Shi et al.,

2016), colour palette, contrast range or the camera re-

sponse function. Other properties, such as an object’s

albedo, texture or shape are often unique to the object

and cannot be derived directly from contextual infor-

mation in the scene. While methods for approximat-

ing these do exist (Gardner et al., 2017), they are dif-

ﬁcult to integrate into end-to-end systems and can be

difﬁcult to parameterise. The recent successes in DL

have motivated a number of approaches (Tsai et al.,

2017; Azadi et al., 2018; Chen and Kae, 2019) which

attempt to exploit the huge amount of natural imagery

available in public datasets in order to learn the map-

ping between a corrupted composite image and a cor-

rected composite, or natural image.

2.2 Multi-task Learning

Due to the abundance of natural image data and the

ill-posed nature of the compositing problem, DL ap-

proaches are well-suited for this task. However, su-

pervised DL methods require large amounts of an-

notated data in order to learn and generalise well.

This requirement grows along with the complexity

of a problem and the desired accuracy. Two popular

DL paradigms, unsupervised learning and multi-task

learning, are often used to address the issues of data

labeling and model generalisation.

In recent years many tasks in image understand-

ing have achieved state-of-the-art performance by in-

corporating multi-task learning Evgeniou and Pontil

(2004), for example in predicting depth and normals

from a single RGB image (Eigen and Fergus, 2015),

detection of face landmarks (Zhang et al., 2014) or

simultaneous image quality and distortion estimation

Kang et al. (2015). This is afforded by the implicit

Towards Unsupervised Image Harmonisation

575

regularization that training a single model for mul-

tiple related tasks imposes (Caruana, 1997) and the

resulting improved generalisation.

State-of-the-art image harmonisation methods fo-

cus largely on improving composites in scenarios

where the identity of pixels belonging to the object

and scene are known a priori. Tsai et al. (2017) use a

DAE-based architecture to map corrupted composites

to corrected ones, incorporating a two-task paradigm,

which attempts to both correct the composite, as well

as segmenting the scene. However, they do not ex-

plicitly condition the network to learn anything more

about the corruption, such as its magnitude, type or

location. Instead they provide location information at

input time, using a binary mask. Chen and Kae (2019)

uses a similar approach - inputting the object mask

at training time, however also introducing mask seg-

mentation and reﬁnement within the GAN, in addition

to geometric transformations. The segmentation net-

work, as part of the adversarial training process, dis-

criminates towards ground truth binary masks as an

output - omitting any perceptual factor in the discrimi-

nation task. This achieves improved results compared

to the DAE, however at the cost of a more complex

architecture and adversarial training.

Due to the many dimensions along which com-

binations of object and scene may vary, compositing

systems should be equipped to assess such differences

before attempting to correct them. Kang et al. (2015)

shows that a multi-task approach is an efﬁcient way

to ensure that distortions are encoded by the model.

3 METHODOLOGY

3.1 Motivation

Whilst multi-task learning has been shown to be efﬁ-

cient in the coupled process of detecting and correct-

ing arbitrary pixel level transformations within im-

ages, perceptually-based encoding of artifacts within

masks has not yet been shown to be effective in the

image harmonisation ﬁeld. Before approaching the

multi-task model, it is necessary to prove empirically

that this end-to-end process is viable. Thus we design

an end-to-end approach using two existing standalone

networks for both detection and harmonisation to test

the efﬁcacy of these perceptual masks in the domain.

3.2 Approach

Our overarching goal is the design of an end-to-end

automatic compositing pipeline, capable of detec-

tion and correction of common compositing artifacts,

Figure 1: Illustration of research methodology adopted in

this work.

without the need for speciﬁcation of an object mask.

In order to evaluate the effectiveness of this approach

we propose to assess predicted perceptually-informed

object masks rather than ground truth object masks as

input to the deep harmonisation algorithm. We then

measure similarity between ground truth images and

composites corrected with the harmonisation algo-

rithm using either the original synthetic binary masks

or the perceptually-based masks predicted by the

detector M

. Accordingly, we refer to composites har-

monised using ground truth masks as C

and compos-

ites generated by the end-to-end system as C

We evaluate the hypothesis that the performance

of an end-to-end detection and harmonisation model

is comparable to a harmonisation model using man-

ually created object masks. Conﬁrmation of this hy-

pothesis would support our case for incorporating ex-

plicit detection of composite artefacts into end-to-

end image composite harmonisation systems. Our re-

search methodology is summarised in Figure 1.

3.3 Detector and Harmoniser Models

Both the detector (Dolhasz et al., 2019) and the har-

moniser (Tsai et al., 2017) are deep, image-to-image,

fully convolutional autoencoder networks. The detec-

tor takes a single image as input and generates a 2-

channel output mask, which encodes probabilities for

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

576

Figure 2: System overview: illustration of the detector and harmoniser combined into an end-to-end composite harmonisation

system. A synthetic composite image is ﬁrst supplied to the detector, which outputs a 2-channel mask indicating detected

negative and positive (not pictured here) luminance shifts. This mask is converted to a single-channel representation by taking

a maximum over predicted pixel-wise probabilities and fed to the harmoniser network. The harmoniser then produces a

harmonised composite, which we compare against the ground truth.

Figure 3: Dataset generation process adapted from Tsai

et al. (2017): a) source image sampled from MSCOCO, b)

corresponding object mask, c) target image, d) target image

object mask, e) result of luminance transfer (Reinhard et al.,

2001) of source - c), to target - e.

each pixel, p, in the input image as being affected by

a negative (channel 0) or a positive (channel 1) per-

ceptually suprathreshold luminance offset. We com-

bine these two suprathreshold channels by taking a

pixel-wise maximum max(p

, p

). This way we gen-

erate a single mask in the same format as M

from

MSCOCO, where each pixel encodes the probability

of a suprathreshold luminance offset. We do not ap-

ply any modiﬁcations to the harmoniser and adopt the

authors’ original trained implementation. The ﬁnal

detector+harmoniser system can be see in Figure 2.

3.4 Dataset

To perform a fair comparison, we follow the compos-

ite generation approach of Tsai et al. (2017). Specif-

ically, we sample pairs of images containing objects

belonging to the same semantic category (e.g. per-

son, dog, bottle etc.) from the MSCOCO dataset (Lin

et al., 2014). Using their corresponding object masks,

we perform statistical colour transfer based on his-

togram matching, proposed by Reinhard et al. (2001).

This process can be see in Figure 3. This colour trans-

fer is performed between object regions of the same

semantic category. As the detector is only conditioned

for luminance offsets, we perform colour transfer only

on the luminance channel of Lab colourspace. We

generate a total of 68128 composites and correspond-

ing ground truth images. We also extract correspond-

ing the ground truth masks for comparison against the

masks predicted by the detector.

3.5 Similarity Metrics

To evaluate each of the two approaches, we calculate

similarity metrics between ground truth images C

and composites corrected by the methods under test:

and C

. We adopt the objective metrics used in the

original work, i.e. Mean Squared Error (MSE):

MSE =

∑

i=0

−

)

(1)

where Y is the ground truth and

Y is the harmonised

image (either C

or C

), and Peak Signal-to-Noise ra-

tio (PSNR):

PSNR = 10log



MSE



(2)

here R is the maximum possible pixel intensity - 255

for an 8 bit image. In addition, we leverage the

Learned Perceptual Image Patch Similarity (LPIPS)

(Zhang et al., 2018), which measures similarity based

on human perceptual characteristics. We denote these

errors with subscripts referring to the method the

composite was ﬁxed with, e.g. MSE

for MSE be-

tween the ground truth image and corresponding com-

posite ﬁxed using predicted masks; MSE

for MSE

between ground truth and a composite ﬁxed using the

original MSCOCO masks.

3.6 Procedure

Using our generated composite dataset we ﬁrst eval-

uate the harmoniser with ground truth masks. We

then use the same dataset to generate predicted ob-

ject masks using the detector and feed these along

with the corresponding composite images to the har-

moniser. We obtain two sets of corrected composites:

composites corrected using the ground truth masks C

and composites ﬁxed using masks predicted by the

detector] C

. We then calculate similarity metrics be-

tween the ground truth images used to generate the

Towards Unsupervised Image Harmonisation

577

Figure 4: A comparison of the distributions for both C

(composites corrected with synthetic ground truth masks) and C

(corrected with masks predicted by the detector) with the number of images in each bin for each metric value. This is shown

for: (a) MSE, (b) PSNR and (c) LPIPS. Larger values of MSE and LPIPS indicate poorer performance, whilst this is true for

smaller values of PSNR.

Figure 5: The image-wise error differentials for C

-C

. This is shown for each of the three metrics: (a) MSE, (b) PSNR and (c)

LPIPS. Note, negative values for MSE and LPIPS indicate images for which C

(composites corrected with masks predicted

by the detector) achieves lower error than C

(composites corrected with synthetic ground truth masks). For PSNR, this is true

for positive values.

composites in the ﬁrst place, and each of the two sets

of corrected images C

and C

. These are reported in

the following section.

4 RESULTS

The results of our evaluation can be seen in Figure

4, which shows distributions of each of the similarity

metrics calculated between ground truth images and

composites ﬁxed using C

and C

respectively. Mean

similarity metrics can be seen in Table 1. Overall,

masks predicted by the detector yield higher average

errors across all three metrics compared to the ground

truth masks, however the magnitude of these differ-

ences is small for each of the metrics. Figure 5 shows

distributions of image-wise error differentials for both

techniques.

5 DISCUSSION

Our results indicate that using detected, instead of

ground truth object masks can yield comparable re-

Table 1: Means of similarity metrics for both techniques

evaluated against ground truth: harmoniser, and the detec-

tor+harmoniser. Lower is better for LPIPS and MSE, higher

is better for PSNR.

Metric harmoniser detector + harmoniser

MSE 19.55 22.65

PSNR 35.81 35.18

LPIPS 0.0227 0.0292

sults when performing automatic image composite

harmonisation. Errors obtained using ground truth

masks are on average lower compared to those ob-

tained using predicted masks, however in a number

of cases the situation is reversed. For example, Fig-

ure 6c and d shows cases of the harmoniser over-

compensating, while the detector+harmoniser com-

bination achieves a more natural-looking result. We

stress that these results were obtained with no addi-

tional training.

Further investigation indicates particular scenar-

ios where this occurs. In some cases, the harmoni-

sation algorithm applies an inappropriate correction,

rendering a higher error for C

compared to the un-

harmonised input. Then, if M

does not approximate

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

578

Figure 6: Examples of the harmoniser with ground truth masks over-compensating, and applying colour shifts to compen-

sate a luminance transform, resulting in suboptimal output. From left: a) ground truth, b) input composite, c) output of

detector+harmoniser, d) output of harmoniser with ground truth masks, e) masks predicted by detector, f) ground truth masks.

well, is blank (no detection) or its average inten-

sity is lower than that of M

, the additional error in-

duced by the harmonisation algorithm is minimised,

rendering lower errors for C

. This can be seen in

both images in 6d. This indicates the beneﬁt of a

perceptually motivated approach to mask prediction,

allowing the inﬂuence over the weight of the trans-

formation applied by the harmoniser. We also no-

tice that the deep harmonisation network tends to ap-

ply colour transformations regardless of whether they

are required. In some cases, the perceptually-based

masks mitigate this problem. Images showing exam-

ples of comparable performance of the two methods

can be found in Figure 7. Subﬁgures c and d show the

results of harmonisation using the apporaches under

test and subﬁgures e and f show M

and M

respec-

tively.

Due to the nature of the detector network currently

operating solely on luminance transforms, a further

beneﬁt to the multi-task learning paradigm is the gen-

eralisability to arbitrary pixel level transforms, for ex-

ample colour shifts. The binary masks accepted by

harmoniser networks currently do not separate across

these transforms, they treat them all homogeneously.

A perceptually motivated approach to the predicted

mask can encode, on a feature-by-feature basis, the

perceptual likelihood of harmonisation required. This

is not to say necessarily that deep harmonisation net-

works cannot learn this behaviour, but further sup-

port to encode this non-linearity at the input to the

network and/or by explicit optimisation at the output,

particularly in a multi-task context, would likely ben-

eﬁt performance and improve generalisation (Caru-

ana, 1997).

6 CONCLUSION

These ﬁndings, obtained by combination of off-the-

shelf models, not modiﬁed or re-trained for this spe-

ciﬁc task, indicate that information about location

and magnitude of composite artifacts can be useful in

improving the performance of existing compositing

and harmonisation approaches. Furthermore, our re-

sults show that the requirement for provision of object

masks for such algorithms can be relaxed or removed

entirely by the explicit combination of composite ar-

tifact detection with their correction. This provides

a basis for investigation in future work of joint mod-

elling of both the detection and correction of compos-

ite image artifacts, e.g. under a multi-task learning

paradigm.

REFERENCES

Agarwala, A., Dontcheva, M., Agrawala, M., Drucker, S.,

Colburn, A., Curless, B., Salesin, D., and Cohen, M.

(2004). Interactive digital photomontage. In ACM

Transactions on Graphics (ToG), volume 23, pages

294–302. ACM.

Azadi, S., Pathak, D., Ebrahimi, S., and Darrell, T. (2018).

Compositional gan: Learning conditional image com-

position. arXiv preprint arXiv:1807.07560.

Burt, P. and Adelson, E. (1983a). The laplacian pyramid as

a compact image code. IEEE Transactions on com-

munications, 31(4):532–540.

Burt, P. J. and Adelson, E. H. (1983b). A multiresolution

spline with application to image mosaics. ACM trans-

actions on Graphics, 2(4):217–236.

Caruana, R. (1997). Multitask learning. Machine learning,

28(1):41–75.

Towards Unsupervised Image Harmonisation

579

Figure 7: Comparison of harmonisation outputs from our evaluation. From left to right: a) ground truth, b) input composite,

c) corrected with detector+harmoniser C

, d) corrected with ground truth masks + harmoniser C

, e) Detected masks M

, f)

ground truth masks M

. Masks in colour indicate the raw output of the detector, where the direction of detected luminance

shifts is indicated - red for negative and green for positive shifts.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

580

Chen, B.-C. and Kae, A. (2019). Toward realistic image

compositing with adversarial learning. In Proceed-

ings of the IEEE Conference on Computer Vision and

Pattern Recognition, pages 8415–8424.

Dolhasz, A., Harvey, C., and Williams, I. (2019). Learning

to observe: Approximating human perceptual thresh-

olds for detection of suprathreshold image transforma-

tions. arXiv preprint arXiv:1912.06433.

Dolhasz, A., Williams, I., and Frutos-Pascual, M. (2016).

Measuring observer response to object-scene dispar-

ity in composites. In 2016 IEEE International Sym-

posium on Mixed and Augmented Reality (ISMAR-

Adjunct), pages 13–18. IEEE.

Eigen, D. and Fergus, R. (2015). Predicting depth, surface

normals and semantic labels with a common multi-

scale convolutional architecture. In Proceedings of

the IEEE international conference on computer vi-

sion, pages 2650–2658.

Evgeniou, T. and Pontil, M. (2004). Regularized multi–task

learning. In Proceedings of the tenth ACM SIGKDD

international conference on Knowledge discovery and

data mining, pages 109–117. ACM.

Gardner, M.-A., Sunkavalli, K., Yumer, E., Shen, X., Gam-

baretto, E., Gagn

e, C., and Lalonde, J.-F. (2017).

Learning to predict indoor illumination from a single

image. arXiv preprint arXiv:1704.00090.

Guillemot, C. and Le Meur, O. (2013). Image inpainting:

Overview and recent advances. IEEE signal process-

ing magazine, 31(1):127–144.

Kang, L., Ye, P., Li, Y., and Doermann, D. (2015). Si-

multaneous estimation of image quality and distortion

via multi-task convolutional neural networks. In 2015

IEEE international conference on image processing

(ICIP), pages 2791–2795. IEEE.

Lalonde, J.-F. and Efros, A. A. (2007). Using color compat-

ibility for assessing image realism. In 2007 IEEE 11th

International Conference on Computer Vision, pages

1–8. IEEE.

Levin, A., Lischinski, D., and Weiss, Y. (2004). Col-

orization using optimization. In ACM transactions on

graphics (tog), volume 23, pages 689–694. ACM.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., Doll

ar, P., and Zitnick, C. L. (2014).

Microsoft coco: Common objects in context. In Euro-

pean conference on computer vision, pages 740–755.

Springer.

erez, P., Gangnet, M., and Blake, A. (2003). Poisson im-

age editing. ACM Transactions on graphics (TOG),

22(3):313–318.

Porter, T. and Duff, T. (1984). Compositing digital images.

In ACM Siggraph Computer Graphics, volume 18,

pages 253–259. ACM.

Ranjan, R., Patel, V. M., and Chellappa, R. (2017). Hy-

perface: A deep multi-task learning framework for

face detection, landmark localization, pose estimation,

and gender recognition. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 41(1):121–135.

Reinhard, E., Adhikhmin, M., Gooch, B., and Shirley, P.

(2001). Color transfer between images. IEEE Com-

puter graphics and applications, 21(5):34–41.

Ruder, S. (2017). An overview of multi-task learn-

ing in deep neural networks. arXiv preprint

arXiv:1706.05098.

Shi, W., Loy, C. C., and Tang, X. (2016). Deep specialized

network for illuminant estimation. In European Con-

ference on Computer Vision, pages 371–387. Springer.

Sunkavalli, K., Johnson, M. K., Matusik, W., and Pﬁster,

H. (2010). Multi-scale image harmonization. ACM

Transactions on Graphics (TOG), 29(4):125.

Tsai, Y.-H., Shen, X., Lin, Z., Sunkavalli, K., Lu, X., and

Yang, M.-H. (2017). Deep image harmonization. In

Proceedings of the IEEE Conference on Computer Vi-

sion and Pattern Recognition, pages 3789–3797.

Vincent, P., Larochelle, H., Bengio, Y., and Manzagol, P.-

A. (2008). Extracting and composing robust features

with denoising autoencoders. In Proceedings of the

25th international conference on Machine learning,

pages 1096–1103. ACM.

Wright, S. (2013). Digital compositing for ﬁlm and video.

Routledge.

Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang,

O. (2018). The unreasonable effectiveness of deep

features as a perceptual metric. In CVPR.

Zhang, Z., Luo, P., Loy, C. C., and Tang, X. (2014). Facial

landmark detection by deep multi-task learning. In

European conference on computer vision, pages 94–

108. Springer.

Towards Unsupervised Image Harmonisation

581