The perception of historic events is frequently shaped by speci昀椀c images that have been ascribed an iconic status. These images are widely reproduced and recognised and can therefore be considered memorable. A question that arises given such images is whether the memorability of iconic images is intrinsic or whether it is shaped. In this work we analyse the memorability of iconic images by means of computational techniques that are speci昀椀cally designed to measure the intrinsic memorability of images. To judge whether iconic images are inherently more memorable we establish two baselines based on datasets of diverse imagery and of newspaper imagery. Our 昀椀ndings show that iconic images are not more memorable than modern day newspaper imagery or when compared to a diverse set of everyday images. In fact, by and large many of the iconic images analysed score on the low end of the memorability spectrum. Additionally, we explore the variation in memorability of reproductions of iconic images and 昀椀nd that certain images have been edited resulting in higher memorability scores, but that the images by and large are reproduced with memorability close to the original.
the course of history. Due to this in昀氀uence iconic images have been widely researched with the aim to understand what makes them unique.
Di昀erent aspects of iconic images have been researched. One o昀琀en noted criteria for
iconicity is the importance of the captured historic event. Due to the democratisation of photography,
images of recent historic events are abundant. Thus an image that merely displays a signi昀椀cant
historic event is not su昀케cient to reach iconic status. For an image to become iconic it must be
widely spread in the media1[
Memorability of images is a complex characteristic to predict; Using only the surface features
of the image for predicting memorability is not comprehensive. Automating the prediction
of memorability with Convolutional Neural Networks (CNN) has been a successful solution
to this problem 1[
Besides the advantages of automatizing the process by using a CNN, the usage of a prediction CNN has another advantage: As iconic photographs are images that are widely spread it is likely that a great share of people has already seen said image. Therefore it would be di昀케cult to test the memorability of these images with a memory task. It would be a challenge to guarantee that these people have never seen these images before. Thus the challenge of testing the memorability of iconic photographs context-independent can be solved by using the above-mentioned MemNet. Additionally, the iconicity of some images is determined by the fame of the subject. MemNet is not trained to recognize celebrities or other symbols. In this research, the focus is on what image aspects the memorability in昀氀uenced, without analysis of the context of the images. This makes the usage of MemNet a great 昀椀t. In this paper, iconic photographs are tested on their memorability by the use of the CNN MemNet to explore to what extent the memorability of iconic photographs can be measured.
To investigate this we formulate the following two research questions: RQ 1: How does the memorability of iconic images compare to a baseline? RQ 2: To what extent does the memorability of iconic images di昀er across variations?
There has been little to no computational research on iconic images, but in the following we
will give a brief overview of other research directions. As iconic images are not a bound set,
di昀erent de昀椀nitions and what criteria they should meet have been proposed [
Additionally, other studies have added criteria to be more comprehensive. For example, in
some studies more stress has been placed on the symbolic meaning that people have of the
image [
But as other factors of the reception, like the symbolic meaning or being a part of the
collective memory, of an iconic image are not fully encoded in the image itself and are subjective,
it is yet challenging to study computational19[]. It is yet di昀케cult to study the semantics of
images computationally, as this o昀琀en is can not be conducted from the visual data or metadata.
Therefore the study of the perception of iconic images is o昀琀en limited to interviews or
largescale surveys [
Due to technological advancements capturing and sharing have become more convenient.
Mass media and its consumption lead people to be exposed to countless images every day.
People are exposed to many images daily and some of those are better remembered than
others. Which image people remember and which they forget is related to the memorability of
the image. Di昀erent academic 昀椀elds have an interest in image memorability. For example,
research in cognitive sciences answered how certain activities in the brain correlate to the
memorability of images. Additionally, image memorability has been a subject of research in
the computer vision 昀椀eld in recent years 1[
When diving further into the correlation between memorability and some image aspects,
collected across di昀erent studies, the following observations were made. In some studies, it
appeared that a number of object categories, like people, animals and vehicles were relatively
more memorable 2[
In this paper, memorability will be measured computationally through a score given by MemNet. MemNet is a CNN trained on an annotated dataset called LaMem. LaMem is a large-scale dataset constructed of multiple already existing datasets that were annotated in a perception study with human subjects who performed a memorability task. The datasets used are highly diverse, which leads to LaMem being a diverse dataset totalling 60.000 image1s6][. It contains images of humans, animals, landscapes, and even art, which also includes abstract images. This diversity of image types is displayed in Figur1e.
For the LaMem annotation task a stream of images was shown to participants and a昀琀er a
varying number of distractors the participants had to indicate if they had seen the image before.
Thus when an image was remembered o昀琀en the image would be scored as more memorable.
A昀琀er performing this task on a group of participants a memorability score for each image
could be calculated. This score is the annotation for each image in the LaMem dataset. This
makes LaMem thereby a su昀케cient foundation for a CNN to compute a memorability score
(0 − 1 scalar) for di昀erent types of images [
The following step-by-step procedure, shown in Figu2rerepresents our proposed method that can be used for measuring a memorability across a set of images.
The procedure we propose involves a comparative process between two datasets (A and B). Here, dataset A represents the collection to be studied, whereas dataset B functions as a baseline. The general idea of this approach is to not only measure the memorability scores but also judge whether the scores di昀er from a meaningful baseline. Each dataset is pre-processed by transforming the images to a suitable input resolution for the CNN (typically a low resolution like256 × 256) and normalizing them. Once the dataset are pre-processed the images will be passed through the pre-trained CNN to obtain a per-image memorability score.
The method for comparing between the memorability scores of datasets A and B depends on the sizes of the respective datasets. In the case of small datasets, the analysis will be mostly done by qualitatively looking at the images to draw conclusions based on the speci昀椀c images. When dealing with large-scale datasets, the comparison can be based on a statistical analysis. Whilst our baseline datasets are large-scale, the dataset of iconic images used is too small scale to reliably perform statistical analysis.
Further analysis beyond the comparative analysis will consist of sampling images which are either random or statically interesting, according to their attribute score. Sampling random images will give an indicative perspective on how the data is constructed. This method of analysis is based on Distant Viewing3[], a method for studying large visual corpora, where the main take way is to perform the computational analysis whilst also viewing the corpora. The latter is import because the intuitive semantics that may get lost in a computational analysis will thereby also be studied. As the semantics are important for iconic images, this method suits this paper well. Conclusions drawn from these analyses can then be linked to prior work.
For our dataset of iconic images we use a selection of 26 images that were initially in a
largescale survey on iconic images 1[
To place the memorability scores in context, we compare to two di昀erent baselines. The 昀椀rst
baseline used is the LaMem dataset; This dataset consists of images with di昀erent subjects
and themes, and functions as a diverse baseline representing a ‘general image collectio1n6]’.[
This is an annotated dataset for memorability, but the annotations were not used; Images of
the LaMem dataset were inserted in MemNet to create a memorability score generated with
the same circumstances as the dataset it is compared with. The second baseline used was the
GoodNews dataset, which consists of all kinds of images used in the New York Times from 1818
until 2019 [
0.9120 Survivor of Hutu death camp 0.8290
Abu Ghraib prisoner 0.8098 Mao Zedong
0.7767 Times Square Kiss
0.7610 Assassination of Inejiro Anasuma 0.7248 Burning
Monk 0.7476 Spanish Soldier 0.6982 Raising a Flag over
the Reichstag 0.6796
Che Guevera
0.8894 The Falling
Man 0.8190 Hinden
burg Disaster 0.7940 Mohandas
Ghandi 0.7725 Napalm Girl
0.7588 Vultre and the girl 0.7462 Raising the
flag on Iwo Jima 0.7173 Coup in
Chile 0.6981 Holocaust survivors 0.6449
Migrant Mother
0.8767 Man on the Moon
0.8168 Alan Kurdi
0.7897 Hijacked Airplane
0.7697 Viet Cong Execution
0.7533 Situation
Room 0.7447 Kent State Shootings
0.7151 Tank Man 0.6939 when compared to other media images.
The distribution of the memorability scores of the LaMem and GoodNews dataset is visualised in Figure5. Additionally, the scores of the iconic images are plotted with a scatter plot at the bottom of the Figure. From the distributions we can observe that the GoodNews images are generally more memorable than the LaMem images. Additionally, the LaMem dataset has a wider distribution than the GoodNews dataset. There are fewer images in the GoodNews dataset with a very low memorability score, which might re昀氀ect that all images in GoodNews have been approved by an editor of the paper. It is very likely an editor would select a picture with at least some traits that correlate with memorability. For example, images that are selected by an editor o昀琀en need to have either a clear main object, thus big object size or some aesthetic value. Taking into account scatter plot of the iconic images we observe that iconic images are generally (slightly) less memorably than the images depicted in the media.
To further clarify how the iconic images compare to both datasets we plot the distribution of the iconic images according to the quartiles of both LaMem and GoodNews in Figu6r.eWe can observe that most iconic images fall into the lower quartiles for both datasets, scoring slightly higher when compared to LaMem. As the LaMem dataset is representative of all kinds of images, it can be concluded that these iconic images are slightly less memorable images. When comparing the iconic images to the GoodNews dataset, it stands out that most of the iconic images have a memorability score that overlaps with the 昀椀rst quantile of the GoodNews dataset. The distribution of the iconic images in the quartiles of the GoodNews dataset is signi昀椀cantly shi昀琀ed to the less memorable side. Thus iconic images are less memorable than images depicted in the media.
Until now we have analysed the most canonical versions of the 26 iconic images. However, it is known that edits have been made in prominently published versions of iconic images. To explore whether these edits have been made (implicitly or explicitly) to boost the memorability of iconic images we further analyse the circulations collected by Smits and R2o5s].[For each of the 26 iconic images they used the Google Cloud Vision API to retrieve online circulations, thereby creating a dataset of 900k images. Per image the number of variations di昀ers from the Che Guevara portrait having over 100k variations retrieved to the image of Mao Zedong and the founding of the PRC, having less than 3k variations. Figur7e shows how many images were retrieved for each of the images.
Due to how the variations were collected, in iterations based on previous retrieval results, the di昀erences to the original get progressively larger the deeper we get into the list of variations. The variations of the iconic image vary in di昀erent crops, di昀erent shades of colors and so on. There are also variations where other images and text are added to the iconic image, this can di昀er from a logo of a broadcaster to a book cover where the original image also appears. Furthermore, internet memes, collages and other types of photo-shopped images appear in the dataset. Additionally, the API is not infallible, there are some images in the dataset where the original image does not appear at all. Roughly speaking we observe that images which have been circulated less there are also fewer edits. To control for some of this variation, and to limit the scale of the comparison, we only use the 昀椀rst 10k variations for each iconic image.
The memorability scores for the variations of each iconic image are visualised in Fig7u. re Images to the right on the x-axis should generally have less resemblance with the original image. When viewing the distributions of scores for the variations some patterns can be found. We recognise four groups: (1) distributions with relatively little spread across the variations, (2) distributions that fan outward towards the end of the plot, (3) distributions that have a large spread from start to 昀椀nish, and (4) distributions where we can recognise clear clusters.
Examples of the 昀椀rst group, with little spread, at the Migrant Mother, the Situation Room, and the assassination of Anasuma. Within this group we can recognise two subcategories, as they are either images that have less than 10k variations or they are images where the canonical form is most dominant. Because for the latter subcategory we have only looked at the 昀椀rst 10k images retrieved, it is still very likely that there are also many variations that di昀er more strongly circulating on the internet. But these would only be retrieved further in the dataset than the 昀椀rst 10k images selected. However, it is still noticeable that the 昀椀rst 10k variations have more resemblance with the original iconic image than for other widely circulated iconic images.
For the second group examples have most of their spread at the end of the graph, such as the Burning Monk, the Falling man and the Tank Man images. In these images, it appears that there is de昀椀nitely a big share of images that have a big resemblance with the original picture but variations that di昀er more already appear within the 昀椀rst 10k variations.
In the third group there is a lot of spread from beginning to 昀椀nish, these are images like Gandhi, Raising the 昀氀ag over Iwo Jima and the Image of the Spanish Soldier. These images appear in a lot of di昀erent variations. This can be explained by it being popular photos for web pages, book covers and other types of editing where the context of the image gets altered heavily.
Lastly is the group where we can recognise di昀erent clusters. This group includes images like Napalm Girl, the Hijacked Plane and the portrait of Sharbat Gula. All these images have an alternative popular variation, which leads to a cluster forming in the 昀椀gure. In the following section, the Napalm Girl image will be highlighted to give an example of what these variations look like. Two variations of the Sharbat Gula image and the Hijacked plane are displayed in Figure8 with their memorability score in the caption of each image. The variations that are highlighted in this 昀椀gure are average images from the main clusters. In the Sharbat Gula variations graph, the bigger cluster on the top of the graph is the original image (Fig8uar)eand the cluster that lies under most of the images are depicted in Figu8rbe. In the Hijacked Plane, there is a similar pattern, the graph of the variations has two distinct clusters, one at the top which consists of images like the original image (Figur8ec and the other cluster which has a lower memorability score than the 昀椀rst cluster. This cluster consists of images like in Fig8udr,e which has a wider crop than the original image such that the building on the le昀琀 still remains in the image. (a) The original image with a memorability score of 0.9120 (b) The variation of the original with a recent image placed alongside has a memorability score of 0.8511 (c) The published image with a memorability (d) A variation of the published image with a score of 0.7697 wider crop and a memorability score of 0.7095
A number of di昀erent crops of the Napalm girl image circulate online. The scatter plot which displays the memorability scores and their variations is depicted in Fig9u.rTehe outliers on the upper side of this plot, images with a score above 0.85, are either images heavily photoshopped or images that got in this dataset but in which the Napalm Girl image does not appear. One of the aspects that stands out from this graph is the big cluster of images which are under the trend line and appear from about 4.000 on the x-axis. When sampling these images it appears that these are mainly the Napalm Girl image with a tighter crop than the original image. This is displayed in Figure10. Where the original image is displayed in Figu1r0ea and the tighter crop in Figure10b. The part of the image that is only visible in the wider crop consists of a big part of the sky and a photographer on the right. The tight crop is more focused on the children, which are the main subject of the picture, and thereby take up a larger portion of the image. The tighter crop having a higher memorability score 昀椀ts the notion that images with a bigger object size are more memorable and hints that memorability might be an (implicit) criteria for edits done by photo-editors.
Another frequent variation was the image in Figur1e0c. This image has a signi昀椀cantly higher score than the original picture and was the only variation with a score in this range that still fully depicted the original image. The original image was edited with a red-hued 昀椀lter over the image. That this red-hued image has such a higher memorability score is expected as the redness of an image has been demonstrated to positively a昀ect memorability.
The 昀椀ndings in this paper are based on the result generated by MemNet. Even though MemNet is a validated method of predicting memorability, the actual memorability and the computed (a) The original full photograph with a memora- (b) Widely published tighter crop with a memobility of 0.6767 rability of 0.7652
(c) A variation with a red 昀椀lter and a memorability of 0.8200 memorability score could still di昀er. In conducted research, it became apparent that image memorability and some qualities were correlated, like aesthetics and certain emotions. But even though they are correlated, MemNet is not trained on these qualities; Thus, for example, images that do not have the more common features that positively in昀氀uence the memorability, like a big object size, but are very memorable because they portray a strong negative emotion that also in昀氀uences memorability, possibly do not get a high memorability score from MemNet. This highlights a mismatch between intrinsic memorability and actually being remembered. A clear example of this mismatch is the Napalm Girl image, this image has a memorability score below the average of LaMem. But Napalm Girl displays a scene of terror, that evokes strong negative emotions like sadness and anger. Those strong emotions make this image more memorable and highlight a limitation of (computational) intrinsic memorability methods. Moreover, frequent exposure of less memorable images might also lead to increased remembrance, which might also play a role for this image.
A possible limitation of this work is that the dataset was selected by Dutch researchers. Despite being selected with the aim to represent international iconic images, the images are predominantly known in the Western World. This could in昀氀uence the results, but this might also interact with MemNet. As MemNet is trained on LaMem which consists mainly of images from the Western World, this bias is is matched - whilst it should not in昀氀uence the analysis itself it does limit to what extent we can generalize about the results. Additionally, the dataset of the iconic images is mostly from the 20th century when color photography was not as common. Most of the 26 images are in black and white. When experimenting with the di昀erences in memorability score for the same image in greyscale as in color, it a昀ected the memorability score; While the LaMem dataset averaged a memorability score 0o.7f645 this score dropped to 0.7456 when all images were converted to greyscale. From this we can observe that colour plays a role, but not to the extent that is changes our conclusions. Even with taking these points into account, the observations we made align with existing theories on memorability.
In answering the research question: How does the memorability of iconic images compare to a baseline? It appeared that of the iconic images, the portraits where the face was a big part of the image had a higher memorability score. This con昀椀rmed the research on memorability where a correlation between higher memorability and big object size, and a face as an image object was established. Additionally, the iconic images with small object sizes and small faces had a lower memorability score. The memorability scores for these iconic images align with theories on memorability. Furthermore, it appeared that images that are depicted in the media are generally more memorable than all other images. Few images are depicted in media that have a relatively low memorability score since all images depicted in media are selected by editors. Iconic images are generally slightly less memorable than other images and are on the lower side of the memorability of images in the media.
To answer the second research question: To what extent does the memorability iconic images di昀er across variations? When comparing the spread of the memorability of the variations of the iconic images, there were certain patterns to be found. Some had clear clusters, which were other popular variations of the original image. Looking at examples of di昀erent variations and memorability scores they show that altering an image can in昀氀uence the memorability score. In these examples, it appeared that variations of the images with a red hue or tighter crop were more memorable. This is in line with previous work.
On the whole we can conclude that computational measures of memorability do not fully capture the memorability of iconic images, as many iconic images are remembered much better than what their memorability score would imply. While the reasons for this may be manifold we expect that frequent exposure and strong emotional content play an important role.