Validation of LLM-Generated Object Co-Occurrence Information for Understanding Three-Dimensional Scenes
This study delves into verifying the applicability of object co-occurrence information generated by a large-scale language model (LLM) to enhance a robot’s spatial ability to understand objects in the real world. Co-occurrence information is crucial in enabling robots to perceive and navi...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10786984/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This study delves into verifying the applicability of object co-occurrence information generated by a large-scale language model (LLM) to enhance a robot’s spatial ability to understand objects in the real world. Co-occurrence information is crucial in enabling robots to perceive and navigate their surroundings. LLM can generate object co-occurrence information based on the learned representations acquired from the learning process. However, the challenge lies in determining whether the co-occurrence gleaned from linguistic data can effectively translate to real-world object relationships, a concept yet to be thoroughly examined. After providing category information about a specific situation, this paper compares and evaluates the co-occurrence degree (co-occurrence coefficient) output by gpt-4-turbo-2024-04-09 (GPT-4) against the object pair data from the ScanNet v2 dataset. The results revealed that GPT-4 achieved a high recall of 0.78 across various situation categories annotated by ScanNet v2, although its precision was relatively low at an average of 0.29. The root mean square error of the co-occurrence coefficient was 0.31. While GPT-4 tends to output slightly higher co-occurrence coefficients, it effectively captures the overall co-occurrence patterns observed in the ScanNet v2 dataset. GPT-4 produced co-occurrence information for more objects than those available in ScanNet v2 while covering co-occurrences among objects within ScanNet v2. These results demonstrate that integrating co-occurrence data from different sources could enhance the ability to recognize real-world objects and potentially strengthen robot intelligence. |
|---|---|
| ISSN: | 2169-3536 |