M3DBench: Let's Instruct Large Models with Multi-modal 3D Prompts



Abstractย 

Recently, 3D understanding has become popular to facilitate autonomous agents to perform further decisionmaking. However, existing 3D datasets and methods are often limited to specific tasks. On the other hand, recent progress in Large Language Models (LLMs) and Multimodal Language Models (MLMs) have demonstrated exceptional general language and imagery tasking performance. Therefore, it is interesting to unlock MLM's potential to be 3D generalist for wider tasks. However, current MLMs' research has been less focused on 3D tasks due to a lack of large-scale 3D instruction-following datasets. In this work, we introduce a comprehensive 3D instructionfollowing dataset called M3DBench, which possesses the following characteristics: 1) It supports general multimodal instructions interleaved with text, images, 3D objects, and other visual prompts. 2) It unifies diverse 3D tasks at both region and scene levels, covering a variety of fundamental abilities in real-world 3D environments. 3) It is a large-scale 3D instruction-following dataset with over 320k instruction-response pairs. Furthermore, we establish a new benchmark for assessing the performance of large models in understanding multi-modal 3D prompts. Extensive experiments demonstrate the effectiveness of our dataset and baseline, supporting general 3D-centric tasks, which can inspire future research.


์ตœ๊ทผ์—๋Š” ์ž์œจ ์—์ด์ „ํŠธ๊ฐ€ ๋” ๋งŽ์€ ์˜์‚ฌ๊ฒฐ์ •์„ ์ˆ˜ํ–‰ํ•˜๋„๋ก ๋•๊ธฐ ์œ„ํ•ด 3D ์ดํ•ด๊ฐ€ ์ธ๊ธฐ๋ฅผ ๋Œ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ๊ธฐ์กด์˜ 3D ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋ฐฉ๋ฒ•๋“ค์€ ์ข…์ข… ํŠน์ • ์ž‘์—…์— ํ•œ์ •๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ์ตœ๊ทผ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)๊ณผ ๋‹ค์ค‘๋ชจ๋‹ฌ ์–ธ์–ด ๋ชจ๋ธ(MLM)์˜ ์ง„๋ณด๋Š” ์ผ๋ฐ˜ ์–ธ์–ด ๋ฐ ์ด๋ฏธ์ง€ ์ž‘์—… ์ˆ˜ํ–‰์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ, ๋‹ค์–‘ํ•œ ์ž‘์—…์— ๋Œ€ํ•œ 3D ์ „๋ฌธ๊ฐ€๋กœ์„œ MLM์˜ ์ž ์žฌ๋ ฅ์„ ๊ฐœ๋ฐฉํ•˜๋Š” ๊ฒƒ์ด ํฅ๋ฏธ๋กญ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ํ˜„์žฌ์˜ MLM ์—ฐ๊ตฌ๋Š” ๋Œ€๊ทœ๋ชจ 3D ์ง€์‹œ ๋”ฐ๋ฅด๊ธฐ ๋ฐ์ดํ„ฐ์…‹์ด ๋ถ€์กฑํ•˜์—ฌ 3D ์ž‘์—…์— ๋œ ์ง‘์ค‘๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” M3DBench๋ผ๋Š” ์ข…ํ•ฉ์ ์ธ 3D ์ง€์‹œ ๋”ฐ๋ฅด๊ธฐ ๋ฐ์ดํ„ฐ์…‹์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ฐ์ดํ„ฐ์…‹์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์ง•์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค: 1) ํ…์ŠคํŠธ, ์ด๋ฏธ์ง€, 3D ๊ฐ์ฒด ๋ฐ ๊ธฐํƒ€ ์‹œ๊ฐ์  ํ”„๋กฌํ”„ํŠธ๊ฐ€ ์–ฝํ˜€ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ๋‹ค์ค‘๋ชจ๋‹ฌ ์ง€์‹œ๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. 2) ์ง€์—ญ ๋ฐ ์žฅ๋ฉด ์ˆ˜์ค€์—์„œ ๋‹ค์–‘ํ•œ 3D ์ž‘์—…์„ ํ†ตํ•ฉํ•˜์—ฌ ์‹ค์ œ ์„ธ๊ณ„ 3D ํ™˜๊ฒฝ์—์„œ์˜ ๊ธฐ๋ณธ ๋Šฅ๋ ฅ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค. 3) 32๋งŒ๊ฐœ ์ด์ƒ์˜ ์ง€์‹œ-์‘๋‹ต ์Œ์„ ํฌํ•จํ•˜๋Š” ๋Œ€๊ทœ๋ชจ 3D ์ง€์‹œ ๋”ฐ๋ฅด๊ธฐ ๋ฐ์ดํ„ฐ์…‹์ž…๋‹ˆ๋‹ค. ๋˜ํ•œ, ๋‹ค์ค‘๋ชจ๋‹ฌ 3D ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ดํ•ดํ•˜๋Š” ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฒค์น˜๋งˆํฌ๋ฅผ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ด‘๋ฒ”์œ„ํ•œ ์‹คํ—˜์€ ์šฐ๋ฆฌ์˜ ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋ฒ ์ด์Šค๋ผ์ธ์˜ ํšจ๊ณผ๋ฅผ ์ž…์ฆํ•˜๋ฉฐ, ์ผ๋ฐ˜์ ์ธ 3D ์ค‘์‹ฌ ์ž‘์—…์„ ์ง€์›ํ•˜๋ฉฐ ํ–ฅํ›„ ์—ฐ๊ตฌ์— ์˜๊ฐ์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.



24b0d121e09c28a8699fe8b115ef046ecc4fc3fe21


24b0d121e09c28a8699fe8b115ef046c61f52e4f9e


https://arxiv.org/abs/2312.10763