Mamba-based S²FI grades fruit quality from video, backed by a dataset of 6,460 clips

The authors of preprint arXiv:2610.12868 propose S²FI (Selective Spatio-Temporal Fruit Inspector), an architecture for sorting fruit by quality that looks at the fruit in motion rather than in a single snapshot. The encoder is built on 3D convolutions, and the core of the model is Mamba, which scales linearly with sequence length. The team collected its own dataset of 6 460 video samples from a conveyor and claims that its method outperforms the compared baseline models on a held-out test set. One key caveat: the abstract gives no specific accuracy or latency figures, so readers will have to look for them in the full text of the paper.

The gap between manual sorting and expensive lines

The authors state the problem plainly. Large intelligent sorting systems cost too much for small and medium-sized fruit and vegetable producers. Manual sorting is cheaper but has three drawbacks: low throughput, criteria that vary from person to person, and a heavy labor requirement. Between these two extremes lies an empty niche, and S²FI tries to fill it: a cheap, lightweight deep learning system that a small farm can deploy quickly.

How the model works

The core idea is that a single frame shows the condition of a fruit poorly. A defect may sit on the far side, or a neighboring fruit or a conveyor part may block it in a given frame. So the fruit is rotated, and the model receives a video clip.

The processing pipeline looks like this:

• A 3D convolutional encoder extracts spatio-temporal features from the clip.
• A Mamba backbone, i.e. a state-space model, captures long-range dependencies between frames with linear complexity. Classic self-attention scales quadratically, and on embedded devices that difference is significant.
• A selective mechanism adaptively weights informative frames during rotation. Frames with blur or occlusion get lower weight.

According to the authors, this approach reduces the impact of motion blur and occlusions, the two main enemies of computer vision on a moving conveyor.

Hardware and data

A separate part of the paper covers the hardware platform. It is built around a dual conveyor and a curved guide: the fruit travels along a rail, rotates, and enters the camera's field of view. The design targets household mains power, fast deployment and lightweight embedded inference.

Key parameters from the abstract:

• Architecture: 3D-CNN encoder + Mamba backbone
• Input: video of a rotating fruit
• Dataset: 6 460 video samples from a conveyor
• Evaluation: held-out test set, comparison with baseline models
• Power: household mains
• Target mode: lightweight embedded deployment

The dataset is a significant part of the contribution in its own right. There are few public video datasets for assessing fruit quality on a conveyor, and a custom collection of 6 460 samples provides a reproducible point of comparison, if the authors release it. The abstract says nothing about that.

What to check before getting excited

The claim of the "best balance of accuracy, speed and practicality" sounds convincing, but for now it remains the authors' word. To judge practical value, look for several things in the full text:

• Which embedded device was used to measure speed, and what throughput, in fruits per minute, it achieved.
• Which crops and which types of defects the dataset includes. A model trained on one type of fruit does not always transfer to another.
• How robust the result is to changes in lighting and camera: in real warehouses, the light rarely matches the lab.
• Whether the code, weights and data are released.

For a farmer, the takeaway is simple. If the paper's numbers hold up, a sorting line on household power with a compact compute unit becomes a realistic alternative to a crew of sorters. For engineers, the more interesting point is different: Mamba in an applied task on embedded hardware shows where linear complexity delivers a real gain, not just a nice line in a benchmark table. The next step is independent validation on other crops and in field conditions.