|
|
--- |
|
|
license: apache-2.0 |
|
|
pipeline_tag: image-text-to-text |
|
|
--- |
|
|
|
|
|
<br> |
|
|
<br> |
|
|
|
|
|
# Spatial-LLaVA-7B Model Card |
|
|
|
|
|
**[Github Repo](https://github.com/xi-jiajun/Spatial-LLaVA)** |
|
|
|
|
|
**[🤗 Huggingface Space Demo](https://huggingface.co/spaces/rogerxi/Spatial-LLaVA)** |
|
|
|
|
|
## 🤖 Model details |
|
|
|
|
|
**Model type:** |
|
|
|
|
|
This finetuned LLaVA model is trained from [liuhaotian/llava-pretrain-vicuna-7b-v1.3](https://huggingface.co/liuhaotian/llava-pretrain-vicuna-7b-v1.3) for improving spatial relation reasoning of large multi-modal model. |
|
|
|
|
|
LLaVA is an open-source chatbot trained by fine-tuning LLaMA/Vicuna on GPT-generated multimodal instruction-following data. |
|
|
It is an auto-regressive language model, based on the transformer architecture. |
|
|
|
|
|
## 🎯 Intended use |
|
|
**Primary intended uses:** |
|
|
The primary use of LLaVA is research on large multimodal models and chatbots. |
|
|
|
|
|
**Primary intended users:** |
|
|
The primary intended users of the model are researchers and hobbyists in computer vision, natural language processing, machine learning, and artificial intelligence. |
|
|
|
|
|
## 📚 Training dataset |
|
|
Instruction following training: [rogerxi/LLaVA-Spatial-Instruct-850K](https://huggingface.co/datasets/rogerxi/LLaVA-Spatial-Instruct-850K) |
|
|
|
|
|
## 📊 Evaluation |
|
|
A collection of 10 benchmarks: |
|
|
| Model | VQAv2 | GQA | VizWiz | SQA | TextVQA | POPE | MME | MM-Bench | MM-Bench-cn | MM-Vet | |
|
|
|:-----------------------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:----------:|:--------:|:-----------:|:--------:| |
|
|
| LLaVA-1.5-7b | 78.5 | 62.0 | **50.0** | 66.8 | 58.2 | 85.9 | **1510.7** | 64.3 | 58.3 | 31.1 | |
|
|
| Spatial-LLaVA-7b | **79.7** | **62.7** | 48.7 | **68.7** | **58.5** | **87.2** | 1472.7 | **67.8** | **60.7** | **31.6** | |
|
|
|
|
|
[Spatial-Relation-Eval](https://huggingface.co/datasets/rogerxi/Spatial-Relation-Eval) (built based on [SpatialRGPT-Bench](https://huggingface.co/datasets/a8cheng/SpatialRGPT-Bench)): |
|
|
### Qualitative Spatial Relations |
|
|
|
|
|
| Model | Below/Above | Left/Right | Big/Small | Tall/Short | Wide/Thin | Behind/Front | Avg | |
|
|
|:-----------------------:|:------------:|:-----------:|:----------:|:-----------:|:----------:|:-------------:|:-------------: | |
|
|
| LLaVA-1.5-7b | 53.91 | 53.49 | 45.36 | 40.00 | **50.00** | 51.04 | 48.97 | |
|
|
| LLaVA-1.5-13b | 54.28 | 52.32 | 45.36 | 48.57 | 49.02 | 47.92 | 49.67 | |
|
|
| Spatial-LLaVA-7b | **56.32** | **66.28** | **60.82** | **48.57** | 49.02 | **52.08** | **55.12** | |
|
|
|
|
|
### Quantitative Spatial Relations |
|
|
|
|
|
| Model | Direct Dist (m / ratio) | Horizontal Dist (m / ratio) | Vertical Dist (m / ratio) | Width (m / ratio) | Height (m / ratio) | Direction (° / ratio) | |
|
|
|:-----------------------:|:------------------------:|:----------------------------:|:--------------------------:|:--------------------------:|:--------------------------:|:--------------------------:| |
|
|
| LLaVA-1.5-7b | 12.90 / 1.06 | 10.68 / 2.03 | 20.79 / 0.94 | **24.19 / 0.50** | 14.29 / 5.27 | 10.23 / 58.33 | |
|
|
| LLaVA-1.5-13b | 13.71 / 0.93 | 10.68 / 3.56 | 16.83 / 0.85 | 15.32 / 0.57 | 17.67 / 5.8 | 14.77 / 54.29 | |
|
|
| Spatial-LLaVA-7b | **24.19 / 0.57** | **14.56 / 0.62** | **41.58 / 0.42** | 22.58 / 1.12 | **18.25 / 2.92** | **20.45 / 56.47** | |
|
|
|
|
|
## 🙏 Acknowledgements |
|
|
We thank Liu Haotian et al. for the LLaVA pretrained script, weights and LLaVA-v1.5 mixture dataset; the teams behind CLEVR, TextCaps, VisualMRC and VQAv2 (via “HuggingFaceM4/the_cauldron”); remyxai for OpenSpaces; Anjie Cheng et al. for Spatial-Bench and data pipeline; Google for OpenImages; and Hugging Face for their datasets infrastructure. |
