CocoaDeep: A preliminary study of the performance sensitivity to datasets of Faster RCNN, YOLO and transformer networks for cocoa pod detection
Philippe Borianne, Frédéric Théveny, Llorenç Cabrera-Bosquet, Sabine-Karen Lammoglia · 2026
Read CocoaDeep: A preliminary study of the performance sensitivity to datasets of Faster RCNN, YOLO and transformer networks for cocoa pod detection by Philippe Borianne, Frédéric Théveny, Llorenç Cabrera-Bosquet, Sabine-Karen Lammoglia (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0351657.txt.
Description
Borianne and colleagues, in a 2026 study, race six artificial intelligence models at spotting cocoa pods in photos of farms in Côte d'Ivoire. Counting pods early helps farmers forecast harvests and bargain with buyers. Photos were low-resolution, as farm pictures often are. Models ranged from tiny versions to ones with many times more internal parts, all trained on thousands of hand-labeled pods.
The small, fast models won. Tiny versions of two YOLO designs beat the bigger transformer and classic models, scoring especially well on trees in the foreground. That upset the usual expectation that bigger is better. Reviewing why, the authors found fault in the training data itself: labels were inconsistent, and there were too few images to feed the giant models properly. The honest conclusion is two-sided: small models already work well for pod counting, and cleaner, larger teaching sets are needed before crowning any champion.
Catalog record
| CREATOR | Philippe Borianne, Frédéric Théveny, Llorenç Cabrera-Bosquet, Sabine-Karen Lammoglia |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65328 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |