Code release of our paper MultiDreamer. Check out our paper, and demo website!
Hae Chan Kim*, Hanbee Jang*, Hyewon Lee*
* Denotes equal contribution
Single-view 3D reconstruction is an actively researched field in computer vision. Many existing models utilize pretrained 2D diffusion models to generate novel views from a single-view image. However, these models face challenges with multi-object input images, struggling to extract accurate 3D positional information. In this work, we introduce MultiDreamer, a method for reconstructing a 3D scene comprising multiple objects from a single-view multi-object image. We enhance 3D mesh accuracy by independently generating meshes for individual objects, utilizing inpainting for obscured regions, and optimizing mesh alignment using the depth map. Our pipeline produces robust 3D scenes for arbitrary objects without training process. Experimental analysis results demonstrate that our pipeline consistently captures the shapes and relative positions of multiple objects compared to the baseline model.
We use the following development environment for this project:
- Nvidia RTX 3090 GPU
- Intel Xeon Processor
- Ubuntu 22.04
- CUDA Version 11.7
- cudatoolkit 10.2.89, 11.7.1
- torch 1.10.1, 1.13.1
- torchvision 0.11.2
- Detectron2 0.6
This code is developed using anaconda3 with Python 3.8 and 3.9 (download here), therefore we recommend a similar setup.
You can simply run the following code in the command line to create the development environment:
$ source setup.sh
We provide two sample input images in the data/assets folder. If you want to test with your own example, the images should consist of exactly two objects. To run the demo, you first need to download the pre-trained model checkpoint files for two models, namely SemanticSAM and SyncDreamer, from this Google Drive folder. The path of the downloaded files MUST follow the structure below :
MultiDreamer/models/
│
├─ SemanticSAM/
│ └─ models/
│ └─ swinl_only_sam_many2many.pth
│
└─ SyncDreamer/
└─ ckpt/
├─ syncdreamer-pretrain.ckpt
└─ ViT-L-14.pt
Before you run demo.sh, you shoud ckeck and modify the path of input image and output directory in demo.sh. If you need, make the data/output/ directory.
INPUT_IMAGE="/MultiDreamer/data/assets/giraffe_and_flower/0_input_giraffe_and_flower.png"
OUTPUT_DIR="/MultiDreamer/data/output/giraffe_and_flower/"
Additionally, demo.sh file contain the code to obtain results of SyncDreamer that can be utilized to compare to our model in evaluation section. If you do not need this part, please comment it out :
python generate.py --input $INPUT_IMAGE --output_dir $OUTPUT_DIR --baseline --mesh
and then, you can run :
$ bash demo.sh
We provide 32 examples in this Google Drive folder. In the link, each example folder contains input png file and ground truth glb file. We recommand setting the downloaded folder as data/eval/.
This is the qualitative result presented in our paper.
In the evaluation part, we compared results of MultiDreamer(Ours) and SyncDreamer(Baseline). We measured Chamfer Distance, Volume IoU, and F-Score for quantitative evaluation. The code to obtain results for both models and compute the metrics is in eval/eval.sh. You can run :
$ conda env create -n eval -f ./env/eval.yaml
$ conda activate eval
$ cd eval
$ bash eval.sh
Finally, you can obtain the result like the table below. The values of each metric may differ from the table, as they are computed from randomly sampled vertices in the inferred mesh and the ground truth mesh.
