In this paper, we propose AR Memory VieWer, which recreates personal memory scenes through AR superimposition of past photos. Deep learning-based feature point matching enables the accurate alignment of personal memory photos with the real-world scene for AR superimposition, even when there are significant visual differences between the past photo and the real-world scene. This enables AR Memory VieWer to provide an eEperience in which users can view past scenes such as landscapes in different seasons or moments when a pet was present through their device. We conducted a user study with the prototype system and confirmed its effectiveness as a new method for recalling personal memories.
Memories are the recollections of events and emotions we have eEperienced in the past,
enriching our lives. Past photos play an important role in evoking or recalling old memories.
While the spread of mobile devices has increased opportunities to take photos, it has been
pointed out that many of these photos remain unused, left behind in folders [
Some studies have effectively utilized past photos taken at the same location as the user's
current position to recreate past scenes through AR superimposition [
One approach is to utilize GPS to enable AR superimposition of past photos corresponding to
the user's location information [
In this paper, we propose AR Memory VieWer, which utilizes deep learning-based feature point matching to accurately align and present past photos as AR superimpositions, even when there are significant visual differences between the past photo and the real-world scene. As shown in Figure 1, AR Memory VieWer enables a new memory-recalling eEperience by allowing users to view personal past scenes through their device.
Using AR Memory Viewer
Camera image
Selected past photo
User’s perspective
An overview of the core processing of AR Memory VieWer is shown in Figure 2. Feature point matching is performed between the camera image and all past photos in the folder, and the past photo with the highest number of matched feature points is selected. A projective transformation is performed using the matched feature points from the selected photo, generating an image in which the past photo is aligned with the camera image. A transformation is applied to the generated image to simulate a magic lens from the user's point of view. The result is then displayed on the screen of the mobile device.
The prototype system used a Surface Pro 7 as the mobile device, and a PC equipped with an
NVIDIA GeForce GTX 1070 GPU as the server for performing deep learning-based feature point
matching. SuperPoint [
The processing results of the prototype system are shown in Figure 3. We verified the operation using scenes from different seasons for outdoor locations and scenes with a pet for indoor locations. Sufficient matching between the selected past photo and the camera image was achieved, enabling a magic lens with accurate alignment of the past photo to the real-world scene.
We conducted a comparative eEperiment to investigate the effectiveness of the proposed method. As a baseline method, we employed a manual selection approach in which participants chose a past photo similar to the real-world scene from a folder and presented it as an AR superimposition at the center of the screen. The set of past photos in the folder consisted of 10 images: one photo used for scene reproduction and nine randomly selected photos from Flickr using the keywords “indoor place” and “outdoor place”.
The eEperiment was conducted in July 2025. The participants were 20 students from our
university (5 females; mean age = 22.9, SD = 1.74). Four past photos taken on our university
campus (two indoor locations taken on June 2025, and two outdoor locations taken on December
2023 and April 2025) were used, and participants eEperienced one of the two methods at each
location. The conditions were counterbalanced so that each was eEperienced an equal number
of times across the four locations. After guiding the participants to each location, we provided
instructions on the appropriate camera angle when using AR Memory VieWer. We measured the
time from launching each application to performing AR superimposition of past photos, as well
as the time from the AR superimposition to eEiting the application. After the eEperience,
participants were asked to complete a questionnaire in which the “game” section of the GEQ [
The time required from launching the application to performing AR superimposition was 18.1 seconds on average (±3.65) for the baseline method and 18.56 seconds on average (±4.31) for the proposed method, showing no substantial difference between the two methods. Since only 10 past photos were included in the folder in this eEperiment, no difference was observed; however, the advantage of the proposed method is eEpected to become more evident as the number of photos increases. Regarding the time from the start of AR superimposition to eEiting the application, the baseline method required 27.82 seconds on average (±11.81), whereas the proposed method required 43.13 seconds on average (±19.89), indicating that the proposed method enabled significantly longer application usage (p < 0.001).
NeEt, the results of the GEQ questionnaire are shown in Figure 4. A paired t-test revealed that the proposed method demonstrated superiority in all subscales eEcept for Tension and Challenge. This indicates that, compared to the baseline method, the proposed method enhances the quality of the application eEperience.
Finally, the results of the custom questions are shown in Figure 5. For Q1, ratings of 7 and 6 accounted for 80% of the responses, suggesting that the concept of the proposed method was sufficiently conveyed. For Q2, all responses were either 7 or 6, indicating that the proposed method may be effective as a new way of referring to personal memories. For Q3, the ratings were limited to 7, 6, and 5, implying that the burden of photo selection was reduced and that past photos were continuously aligned with the real-world scene at the correct position.
In this paper, we proposed AR Memory VieWer, which utilizes deep learning-based feature point matching to accurately align personal memory photos with the real-world scene for AR superimposition. AR Memory VieWer allows users to view past scenes through their device even when the appearance of past photos differs from the real-world scene, as long as the geometric structure is similar. The results of the evaluation eEperiment demonstrated that the concept of the prototype system was sufficiently conveyed and that it is effective as a new way of referring to personal memories.
In the prototype system, feature point matching with SuperGlue was difficult to perform
solely on a mobile device, so communication with a desktop PC was used. However, LiteGlue [
During the preparation of this work, the authors used GPT-5 for grammar and language refinement. After using this tool, the authors reviewed and edited the content as needed and took full responsibility for the publication’s content.