Cross-Generation Optimization of YOLOv26, YOLOv11, and YOLOv8 for Fine-Grained Small-Object Detection and Instance Segmentation in Complex Orchards

In the quiet, sun-dappled rows of a commercial apple orchard, a subtle but difficult challenge awaits the machines of the future. For decades, computer vision has learned to recognize objects, finding cars in traffic or people in crowds with remarkable speed. But when those same machines look at a young apple tree in early summer, the task becomes surprisingly hard. The fruit is tiny, often no larger than a thumbnail, and it is hidden within a dense, chaotic tangle of green leaves, stems, and branches that look almost identical to the fruit itself. This "green-on-green" environment creates a visual puzzle where the target blends seamlessly into the background. To build robots that can care for these trees—perhaps to thin out excess fruit so the remaining ones grow large and sweet—engineers need more than just a camera that sees a fruit. They need a system that can distinguish the delicate, thread-like stem holding the fruit, the small cup at its base, and the fruit body itself, even when they are partially hidden or only a few pixels wide on a screen.

This is the specific problem tackled by a team of researchers at Cornell University, who set out to test how well modern artificial intelligence handles these tiny, hidden details. They focused on a family of AI models known as YOLO, which are famous for their ability to spot objects in real-time. The team did not invent a new type of AI; instead, they acted as careful experimenters, taking three different generations of these models—the established YOLOv8, the newer YOLOv11, and the very recent YOLOv26—and testing them under a wide variety of conditions. They wanted to know two things: does a bigger, more complex AI model always work better for these tiny targets, and does feeding the AI a sharper, higher-resolution picture help it see what it might otherwise miss? To find out, they trained thirty different versions of these models on a dataset of 600 images taken from a real apple orchard in Washington state, where the fruit was at its smallest and most difficult to spot.

The results of this extensive testing revealed a counterintuitive truth about how these machines learn to see. For a long time, the assumption in the field has been that if you want better performance, you simply need a bigger brain. In the world of AI, this means using a model with more layers and more parameters, essentially giving the computer more memory and processing power to analyze the image. The researchers tested this by comparing the smallest versions of the models, known as "nano," against the massive "extra-large" versions. They found that simply making the model bigger did not guarantee better results. In fact, the largest, most computationally heavy models often failed to outperform their much smaller, more efficient cousins. The biggest models sometimes missed the tiny stems entirely or struggled to draw the precise boundaries of the fruit, despite having far more computing power at their disposal.

Instead, the key to seeing these small objects clearly lay in how the images were presented to the AI during training. The researchers compared two approaches: one where the images were shrunk down to a standard size of 640 by 640 pixels, and another where the images were kept at a higher resolution of 960 by 960 pixels. When the images were shrunk, the tiny stems and small fruit bodies lost detail, becoming blurry or disappearing into the noise of the background leaves. By keeping the resolution higher, the researchers preserved the fine spatial details that the AI needed to make sense of the scene. This higher-resolution approach consistently helped the models, but it did not help all models equally. The most significant improvements were seen in the small-to-medium sized models. These compact models, when trained on the sharper images, were able to detect and outline the tiny fruit parts with a level of accuracy that the massive models could not match, even when the massive models were given the same high-resolution pictures.

The study highlighted that the difficulty of the task varies depending on which part of the fruit the robot needs to see. The main body of the young fruit, which is rounder and larger, was relatively easy for the AI to find. The flower cup at the base of the fruit was harder, but still manageable. The most difficult target was the peduncle, the thin, needle-like stem that attaches the fruit to the branch. This structure is so narrow and often so obscured by leaves that it is easily lost when an image is resized or when the AI model is not tuned correctly. The researchers found that the small models trained on high-resolution images were the most successful at finding these stems. For instance, one of the small models, when trained on the 960-pixel images, achieved a high level of accuracy in identifying the fruit and its parts while using only a fraction of the computing power required by the largest models. The largest models, by contrast, required significantly more time to process each image and consumed much more energy, yet they did not produce better results. In some cases, the largest models actually performed worse, suggesting that adding more complexity can sometimes confuse the AI rather than help it.

This work offers a clear guide for engineers building robots for agriculture. It suggests that the path to better perception is not necessarily to build bigger, more expensive computers, but to be smarter about how the data is prepared. By focusing on preserving the fine details of the image and pairing that with a model that is just large enough to handle the task, it is possible to achieve high accuracy without the heavy computational cost. The researchers found that a small model trained on high-resolution images could identify the fruit and its delicate parts with an accuracy that rivals or exceeds that of the most powerful systems available. This is a crucial finding for the future of robotic farming, where robots need to operate quickly and efficiently in the field, often with limited power and computing resources. The study concludes that for the specific challenge of seeing tiny, hidden objects in a complex green environment, the combination of a compact model and a sharp view is far more effective than simply throwing more computing power at the problem. The implementation of these models and the data used are now available for others to use, providing a practical foundation for the next generation of agricultural robots that can see the orchard not just as a green blur, but as a collection of individual, delicate structures waiting to be tended.