DivScene: Towards Open-Vocabulary Object Navigation with Large Vision Language Models in Diverse Scenes

์ €์ž: Zhaowei Wang, Hongming Zhang, Tianqing Fang, Ye Tian, Yue Yang, Kaixin Ma, Xiaoman Pan, Yangqiu Song, Dong Yu | ๋‚ ์งœ: 2024-10-03 | URL: https://arxiv.org/abs/2410.02730 📄 PDF


Essence

Figure 2

Figure 2: Data collection process. On the left, we show the process of collecting scenes. We prompt GPT-4o to

Large Vision-Language Models (LVLMs)์˜ embodied ํ™˜๊ฒฝ ์ดํ•ด์™€ ๋„ค๋น„๊ฒŒ์ด์…˜ ๋Šฅ๋ ฅ์„ ํƒ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด 81๊ฐœ ์žฅ๋ฉด ์œ ํ˜•๊ณผ 5,707๊ฐœ ๊ฐ์ฒด ๋ฒ”์ฃผ๋ฅผ ํฌํ•จํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ์…‹ DivScene์„ ์ œ์‹œํ•˜๊ณ , CoT ์„ค๋ช…์„ ํ†ตํ•œ fine-tuning์œผ๋กœ GPT-4o๋ฅผ 20% ์ด์ƒ ์ƒํšŒํ•˜๋Š” ์„ฑ๋Šฅ ๋‹ฌ์„ฑ.

Motivation

Achievement

Figure 2

Figure 2: Data collection process. On the left, we show the process of collecting scenes. We prompt GPT-4o to

How

Figure 2

Figure 2: Data collection process. On the left, we show the process of collecting scenes. We prompt GPT-4o to

Originality

Limitation & Further Study

Evaluation

Novelty: 4/5 Technical Soundness: 3/5 Significance: 4/5 Clarity: 4/5 Overall: 4/5

์ดํ‰: ์ด ๋…ผ๋ฌธ์€ open-vocabulary object navigation ์ž‘์—…์„ ์ฒ˜์Œ ์ฒด๊ณ„์ ์œผ๋กœ ์ •์˜ํ•˜๊ณ  ๊ธฐ์กด์˜ 100๋ฐฐ ์ด์ƒ ๋‹ค์–‘ํ•œ ๊ฐ์ฒด๋ฅผ ํฌํ•จํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๋ฒค์น˜๋งˆํฌ๋ฅผ ์ œ์‹œํ•˜์—ฌ ๋†’์€ ํ•™์ˆ ์  ๊ธฐ์—ฌ๋„๋ฅผ ๊ฐ€์ง. LVLM์˜ embodied AI ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ค‘์š”ํ•œ ์ž์‚ฐ์„ ์ œ๊ณตํ•˜๋ฉฐ, BFS ๊ธฐ๋ฐ˜ ์ด๋ชจํ…Œ์ด์…˜ ๋Ÿฌ๋‹๊ณผ CoT ์„ค๋ช…์˜ ์กฐํ•ฉ์œผ๋กœ ์‹ค์šฉ์ ์ด๊ณ  ํšจ์œจ์ ์ธ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•œ ์ ์ด ํƒ์›”ํ•จ.

← ๋ชฉ๋ก์œผ๋กœ ๋Œ์•„๊ฐ€๊ธฐ

๐ŸŽง Audio Overview

์ด ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ํŒŸ์บ์ŠคํŠธํ˜• ์˜ค๋””์˜ค๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. (Gemini ยท ํ‚ค๋Š” ๋ธŒ๋ผ์šฐ์ €์—๋งŒ ์ €์žฅ ยท ์™„์„ฑ๋ณธ์€ ์ด๋ฉ”์ผ๋กœ๋„ ์ „์†ก)
โ–ธ ๊ณ ๊ธ‰: ๊ตฌ์„ฑ ๋ฐฉํ–ฅ(๋Œ€๋ณธ ์ž‘์„ฑ ์ง€์นจ) ์ง์ ‘ ์ˆ˜์ •