From Aerospace Dreams to NASA Challenge Winner: Motoki Kimura

The NASA Crater Detection Challenge brought together talented members of the Topcoder community to solve a problem inspired by real planetary science: in the pursuit of next generation, terrain-based optical navigation, NASA is developing a system that will use a visible-light camera on a spacecraft to capture orbital images of lunar terrain and process the imagery to aid in autonomous navigation in space. Rising to the top of the leaderboard was Motoki Kimura / [motokimura], a computer vision research engineer from Tokyo, who earned 1st place with an impressive and highly optimized solution.

Motoki’s journey is a great example of the talent and determination within the Topcoder community. With nearly ten years of experience in computer vision, he has built systems for real-world applications, particularly focusing on models that must run efficiently in resource-constrained environments. Even more impressive, he joined the challenge only a few weeks before the deadline, relying on smart engineering decisions, open-source tools, and his deep technical expertise to quickly develop a competitive solution.

For many Topcoder members, challenges like this are more than competitions—they’re opportunities to test ideas, learn from other experts around the world, and push the limits of what’s possible. In this interview, Motoki shares his journey into computer vision, what motivates him to compete, and how he approached building a winning solution for NASA.

Let’s dive into the conversation.

Tell me a bit about yourself

I’m a computer vision research engineer based in Tokyo. My work focuses on developing and deploying computer vision models for real-world applications, with a particular interest in optimization for resource-constrained environments. Alongside my professional work, I actively participate in international computer vision competitions, including the NASA Crater Detection Challenge and several SpaceNet Challenges hosted on Topcoder.

Outside of coding and AI, what are some hobbies or interests people might be surprised to learn about you?

Outside of coding, I enjoy exploring designer pieces like clothing and eyewear from different countries, particularly from emerging but lesser-known designers who focus on craftsmanship and thoughtful design details. I also enjoy specialty coffee and am always on the lookout for great beans, so quite a bit of my Topcoder prize money tends to disappear into coffee beans.

How long have you been working in the computer vision space and how did you get started?

I’ve been working in computer vision for about ten years. I got started around the time deep learning was becoming widely adopted in the field, when models like YOLO were first introduced. In my first job after graduating, I worked on analyzing satellite imagery using deep learning. That experience got me interested in computer vision, especially applications for remote sensing, and eventually led me to participate in remote-sensing challenges such as SpaceNet and the NASA Crater Detection Challenge.

What motivates you to take part in challenges like this?

Several things motivate me to take part in challenges like this. One is the excitement of competition. Watching your position move on the leaderboard and waiting for the final ranking always brings a unique kind of tension. I also enjoy competing with strong participants from around the world, which pushes me to seriously test both my ideas and my implementation skills. Reviewing other competitors’ solutions after the challenge is also a great way to learn. Since I enjoy developing practical computer vision models in my professional work, these challenges are one of the best ways for me to improve as an engineer.

What was your overall strategy to win this challenge?

My strategy was to take a practical and reliable approach. I joined the challenge relatively late, about three weeks before it ended, so I focused on methods that could produce strong results in a limited time.

Instead of implementing completely new ideas from scratch, I relied heavily on existing methods and open-source tools. For example, I used deep learning-based instance segmentation to detect craters as ellipse-shaped masks. Building such systems from scratch would require significant effort, but many excellent open-source frameworks already exist for training and efficient inference. By combining these tools effectively, I was able to build a strong solution in a relatively short time.

In my professional work, I also frequently rely on open-source tools when developing computer vision models. That experience helped me move quickly in this challenge and achieve a good result despite joining late.

How did you deal with the crater filtering rules (size thresholds, visibility constraints)?

Very small craters, very large craters, and craters partially outside the image boundaries were excluded from evaluation, and detecting them resulted in penalties. My approach was quite simple. I excluded such craters from the training targets so that the model would not detect them. In the few cases where such craters were still detected, I removed them using simple rule-based filtering as a post-processing step.

What part of the challenge did you enjoy the most?

What I enjoyed most was the strict constraints on runtime and peak RAM usage. Because of these constraints, there was a lot of room for optimization and engineering creativity. For example, I selected a lightweight instance segmentation model, used an inference runtime optimized for ARM CPUs, and applied techniques such as image tiling and downscaling before feeding them to the model. I also moved a heavy operation out of the model and reimplemented it as an efficient post-processing step.

What made this challenge particularly difficult for you?

A key challenge was ensuring that the solution remained reliable while applying various optimizations. For example, when converting the model for an ARM-optimized inference runtime, I needed to confirm that the predictions remained consistent before and after conversion. Similarly, when moving a heavy operation out of the model and reimplementing it as a post-processing step, I spent time ensuring that the behavior remained correct.

Can you give a more detailed explanation of your solution?

My solution follows the pipeline illustrated in the figure below.

First, each 2592×2048 image is split into overlapping tiles so that inference can be performed efficiently under the runtime and memory constraints. Each tile is downscaled and then fed into a lightweight instance segmentation model (RTMDet-Ins-s from MMDetection) converted to OpenVINO for efficient inference on ARM CPUs.

The predicted instance masks are then converted into ellipse parameters using OpenCV’s fitEllipse function. After projecting the detections back to the original image coordinates, detections from overlapping tiles are merged and deduplicated using non-maximum suppression.

Finally, detections are filtered and the top 11 craters are selected based on confidence scores to maximize the challenge’s evaluation metric.

In the final evaluation environment (AWS m8g.xlarge, single-threaded), the pipeline achieves a CDA score of 93.34, with approximately 450 MB peak RAM usage and a runtime of about 3.7 seconds per image. 

What does winning a NASA challenge on Topcoder mean to you, both professionally and personally?

Winning a NASA challenge on Topcoder means a lot to me, both professionally and personally.

From a personal perspective, I studied aerospace engineering at university and have long had a strong interest in space and planetary science. Although my career later shifted toward computer vision and AI, participating in a challenge related to NASA’s planetary exploration and applying my work to that field makes the experience especially meaningful to me.

Professionally, the challenge was also very rewarding. It required not only building an accurate model but also engineering a solution that met strict runtime and memory constraints. Solving this kind of practical problem is something I truly enjoy in my work, so it was very satisfying to see the approach perform well among strong competitors on the Topcoder platform.