Lumen: Unleashing Versatile Vision-Centric Capabilities of Large Multimodal Models

Lumen: Large multimodal model with versatile vision-centric capabilities enhancement

News

• [2024.9.26] Lumen is accepted to NeurIPS2024 🎉🎉🎉
• [2024.5.30] Upgraded version of Lumen-v1.5 is introduced. Lumen-v1.5 extends versatile vision-centric capabilities while maintaining general-purpose conversational capabilities.
• [2024.3.13] Lumen paper is released on the arxiv.
• [2024.3.12] Lumen GitHub repo is created.

Lumen: Unleashing Versatile Vision-centric Capabilities of [Paper]
Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, Yu-Gang Jiang

Abstract

Large Multimodal Model (LMM) is a hot research topic in the computer vision area and has also demonstrated remarkable potential across multiple disciplinary fields. A recent trend is to further extend and enhance the perception capabilities of LMMs. The current methods follow the paradigm of adapting the visual task outputs to the format of the language model, which is the main component of a LMM. This adaptation leads to convenient development of such LMMs with minimal modifications, however, it overlooks the intrinsic characteristics of diverse visual tasks and hinders the learning of perception capabilities. To address this issue, we propose a novel LMM architecture named Lumen, a Large multimodal model with versatile vision-centric capability enhancement. We decouple the LMM's learning of perception capabilities into task-agnostic and task-specific stages. Lumen first promotes fine-grained vision-language concept alignment, which is the fundamental capability for various visual tasks. Thus the output of the task-agnostic stage is a shared representation for all the tasks we address in this paper. Then the task-specific decoding is carried out by flexibly routing the shared representation to lightweight task decoders with negligible training efforts.

Performances

Object Detection

Type	Model	Input Size	mAP	AP50	AP75
Specialists	Faster R-CNN-R50	1333*800	40.3	61.0	44.0
	DETR-DC5	1333*800	43.3	63.1	45.9
Vision Generalists	Pix2Seq-v2	1024*1024	46.5	-	-
	UniPerceiver-v2	1600*1400	58.6	-	-
LMM Generalists	Griffon-13B	448*448	24.8	40.6	25.1
	Lumen-7B	448*448	33.9	51.2	34.2
	Lumen-7B-v1.5	448*448	35.3	53.2	35.8

Instance Segmentation

Type	Model	mAP	AP50	AP75
Specialists	Mask R-CNN-R50	37.1	58.4	40.1
	PolarMask	30.5	52.0	31.1
Vision Generalists	Pix2Seq-v2	38.2	-	-
	UniPerceiver-v2	50.6	-	-
LMM Generalists	Lumen-7B	29.1	47.5	29.6
	Lumen-7B-v1.5	30.4	49.8	31.0

Pose Estimation

Type	Model	mAP	AP50	AP75
Specialists	CPM	62.7	86.2	70.9
	RTMPose	68.2	88.3	75.9
Vision Generalists	Pix2Seq-v2	64.8	-	-
LMM Generalists	Lumen-7B	65.4	90.4	72.2
	Lumen-7B-v1.5	67.2	90.4	75.6

VQA

Model	EN_MMBench_DEV	SEED_IMG	MME	MMMU_VAL	MathVista
InstructBLIP	36.0	58.8	1213/292	32.9	25.3
MiniGPT-4	24.3	47.4	582/144	-	23.1
Shikra	58.8	-	-	-	-
Qwen-VL-Chat	60.6	58.2	1488/361	35.9	-
LLaVA-v1.5	64.3	66.1	1511/296	35.6	23.5
Lumen-7B-v1.5	64.9	65.8	1426/332	35.2	24.6

Citation

If you find this project useful in your research, please consider citing:

@article{jiao2024lumen,
  title={Lumen: Unleashing Versatile Vision-Centric Capabilities of Large Multimodal Models},
  author={Jiao, Yang and Chen, Shaoxiang and Jie, Zequn and Chen, Jingjing and Ma, Lin and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2403.07304},
  year={2024}
}

Acknowledgement

• This work is built upon the LLaVA, LISA and SAM.