Compass: Degradation-Simulated Reciprocal Learning with Lightweight Needle RWKV for Multimodal Crack Segmentation under Missing Modalities
This paper introduces Compass, a lightweight multimodal crack segmentation framework that employs Degradation Simulation Distillation, a Needle RWKV backbone, and Evidential Topology-Preserving Fusion to achieve state-of-the-art performance under arbitrary missing modality conditions while maintaining low computational cost.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a detective trying to solve a mystery, but your team is missing some crucial clues. Maybe the camera lens is cracked, the microphone is dead, or the thermal sensor is fogged up. In the world of artificial intelligence, this is a common problem called "missing modalities." AI systems often rely on multiple types of data—like a regular photo (RGB), a heat map (infrared), or a 3D depth scan—to understand the world. When one of these data streams disappears or gets corrupted, many AI systems get confused and fail to do their job. This is a huge deal for safety, especially when we are talking about finding tiny cracks in bridges, roads, or buildings. If an AI misses a crack because a sensor failed, the consequences could be dangerous. The goal, then, is to build an AI that is as tough as a veteran detective: one that can solve the case even when half its evidence is missing, and one that doesn't need a supercomputer to do it.
This is where a new system called Compass comes in. Think of Compass as a super-smart, lightweight robot designed specifically to find cracks in industrial structures, even when its sensors are acting up. The researchers behind Compass realized that most AI models are like students who only study for a test when they have all their textbooks. If you take away half the books, they panic. Compass, however, is trained differently. It uses a clever trick called "Degradation-Simulated Reciprocal Learning." Imagine a student who practices for a test by intentionally throwing away their notes and trying to solve the problems from memory, while a friend who has the notes helps them out. They swap roles, teaching each other how to handle the missing information. This way, when the real test comes and a sensor fails, the AI doesn't freeze; it knows exactly how to fill in the gaps.
The paper introduces Compass as a "lightweight" network, meaning it's small and fast enough to run on portable devices, like the computers inside a drone or a self-driving car. It uses a special backbone called "Needle RWKV." If you imagine a standard AI trying to see a crack as looking at a whole room at once, Needle is like a needle-threading machine that focuses specifically on the long, thin, winding paths of the cracks, ignoring the rest of the background. It's designed to understand that cracks have a direction and a shape, and it follows them like a thread.
To glue all the different pieces of information together, Compass uses something called "Evidential Topology-Preserving Fusion." Think of this as a team meeting where every sensor reports what it sees, but they also rate how confident they are. If the depth sensor is foggy (missing data), it says, "I'm not sure," and the system listens more to the clear infrared camera. But it doesn't just pick the loudest voice; it uses a mathematical rule (Dempster-Shafer theory) to combine their opinions in a way that keeps the crack looking like a continuous line, not a broken-up mess.
The results are quite impressive. The researchers tested Compass on three different datasets containing images of cracks. In one extreme test, they removed 90% of the depth data (basically making the 3D sensor almost useless). Even with this massive handicap, Compass still managed to find the cracks with an F1 score of 0.8216 and a mean Intersection over Union (mIoU) of 0.8434. To put that in perspective, other top methods struggled significantly under the same conditions, often breaking the cracks into tiny, unrecognizable pieces. Compass did all this with only 2.58 million parameters, making it incredibly efficient compared to other models that are hundreds of times larger.
The paper explicitly argues against methods that try to "generate" missing data from scratch or use massive, heavy models that require huge servers. The authors suggest that these approaches are too slow for real-world use and often fail when the missing data is too severe. Instead, they show that by simulating the worst-case scenarios during training and using a smart, direction-aware backbone, you can build a system that is both robust and fast. They didn't just guess this works; they measured it extensively across different datasets and missing ratios, showing that Compass consistently outperforms the current state-of-the-art methods while using a fraction of the computing power. It's a step toward making AI that can actually survive in the messy, imperfect real world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.