Contents

• TokenFusion Description
• Model Architecture
• Dataset
• Environment Requirements
• Script Description
  • Script and Sample Code
    • Training Process
    • Evaluation Process
• Model Description
  • Performance
    • Evaluation Performance
• Description of Random Situation
• ModelZoo Homepage

TokenFusion Description

TokenFusion is a multimodal token fusion method tailored for transformer-based vision tasks. To effectively fuse multiple modalities, TokenFusion dynamically detects uninformative tokens and substitutes these tokens with projected and aggregated inter-modal features. Residual positional alignment is also adopted to enable explicit utilization of the inter-modal alignments after fusion. The design of TokenFusion allows the transformer to learn correlations among multimodal features, while the single-modal transformer architecture remains largely intact. Extensive experiments are conducted on a variety of homogeneous and heterogeneous modalities and demonstrate that TokenFusion surpasses state-of-the-art methods in three typical vision tasks: multimodal image-to-image translation, RGB-depth semantic segmentation, and 3D object detection with point cloud and images.

Paper: Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, Yunhe Wang. Multimodal Token Fusion for Vision Transformers. In CVPR 2022.

Model architecture

The overall architecture of TokenFusion is show below:

Dataset

Dataset used: NYUDv2

• Dataset size：colorful images and depth images, with labels in 40 segmentation classes
  • Train：795 samples
  • Test：654 samples
• Data format：image files
  • Note：Data will be processed in utils/datasets.py

Environment Requirements

• Hardware (GPU)
• Framework
  • MindSpore
• For more information, please check the resources below:
  • MindSpore Tutorials
  • MindSpore Python API

Script Description

Script and Sample Code

.TokenFusion
├── README.md               # descriptions about TokenFusion
├── models
│   ├── mix_transformer.py  # definition of backbone model
│   ├── segformer.py        # definition of segmentation model
│   └── modules.py          # TokenFusion operations
├── utils
│   ├── datasets.py         # data loader
│   ├── helpers.py          # utility functions
│   ├── transforms.py       # data preprocessing functions
│   └── meter.py            # utility functions
├── eval.py                 # evaluation interface
├── cfg.py                  # configure file
├── config.py               # configure file

Training process

To Be Done

Evaluation Process

Launch

# infer example

python eval.py --checkpoint_path  [CHECKPOINT_PATH]

Checkpoint can be downloaded at here or Mindspore Hub.

Result

result: IoU=54.8, ckpt= ./tokenfusion_ascend_v180_nyudv2_research_cv_acc54.8.ckpt

Description of Random Situation

We set the seed inside datasets.py.

ModelZoo Homepage

Please check the official homepage.