Cross-domain zero-shot semantic segmentation for unstructured environments via EVA-CLIP model, ensemble prompt engineering, and optimized text-image matching
Nana Zhou, Xianhua Zhao, Fengjun Zhou, Ji Li, Xiuming Yue · 2026
Read Cross-domain zero-shot semantic segmentation for unstructured environments via EVA-CLIP model, ensemble prompt engineering, and optimized text-image matching by Nana Zhou, Xianhua Zhao, Fengjun Zhou, Ji Li, Xiuming Yue (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-0352325.txt.
Description
Nana Zhou and four colleagues teach ground robots to label scenes they have never been trained on. Their 2026 study targets unstructured environments: rubble, scrub, disaster ground. Ordinary segmentation models need retraining for each new setting. Their framework instead leans on a vision-language model that already links images to words. Deep prompt tuning nudges the image encoder toward rough terrain. The zero-shot ability of the base model survives the tuning.
An ensemble of text prompts, tailored to wild settings, sharpens the matching. Global and local features from both encoders are combined to align words with the right pixels. Vehicles that can segment an unseen hillside into rock, brush, and path without new training data are far more deployable. The design aims squarely at rescue and exploration robots, which never know what terrain waits around the next bend.
Catalog record
| CREATOR | Nana Zhou, Xianhua Zhao, Fengjun Zhou, Ji Li, Xiuming Yue |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53919 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |