* The training process of PP-Matting model uses one GPU and batch size 4.
### 1.2 Datasets
* The common object matting dataset is Compositon-1k or Distinctins-646 (the use of both datasets are requested from the author). COCO2017 and Pascal VOC 2012 are used as the background datasets.
* Human matting uses the private dataset.
### 1.3 Benchmark
|Model | Description | Input shape |
|---|---|---|
|ppmatting_hrnet_w48 | Common object matting | 512 |
|ppmatting_hrnet_w18 | Human matting | 512 |
## 2. Inference Benchmark
### 2.1 Environment
* The PP-Matting model's inference speed test is tested with one V100, batch size=1, CUDA 10.2, CUDNN 7.6.5, PaddlePaddle-gpu 2.3.2.
### 2.2 Datasets
* Common object matting: the test dataset of Compositon-1k or Distinctions-646.
* Human matting: the human image of PPM-100 and AIM-500, total 195 images, named PPM-AIM-195.
### 2.3 Benchmark
| Model | Dataset | SAD | MSE | Grad | Conn |Params(M) | FLOPs(G) | FPS |
"In many image matting algorithms, in order to pursue precision, trimap is often provided as auxiliary information, but this greatly limits the application of the algorithm. PP-Matting, as a trimap-free image matting method, overcomes the disadvantages of auxiliary information and achieves SOTA performance in Composition-1k and Distinctions-646 datasets. PP-Matting uses Semantic Context Branch (SCB) to extract high-level semantic information of images and gradually guides high-resolution detail branch (HRDB) to extract details in transition area through Guidance Flow. Finally, alpha matte is obtained by fusing semantic map and detail map with fusion module.\n",
"\n",
"More details can be found in the paper: https://arxiv.org/abs/2204.09433.\n",
"\n",
"More about PaddleMatting，you can click https://github.com/PaddlePaddle/PaddleSeg/tree/develop/Matting to learn.\n",
"\n"
]
},
{
"cell_type": "markdown",
"id": "8cbcf510-dc56-43f9-9864-5e1de3c7b272",
"metadata": {},
"source": [
"## 2. Model Effects and Application Scenarios\n",
"The human matting effects of PP-Matting are as follows:\n",
"* Clear semantic prediction and detail prediction tasks by branch design.\n",
"* Semantic Context Branch(SCB) ensures the accuracy of the overall prediction of the image by semantic prediction, which roughly divides the image into three parts: foreground, background and transition area.\n",
"* High-Resolution Detail Branch (HRDB) keep high resolution while extracting features，which ensures that the details are not lost.\n",
"* The design of Guidance Flow enables the HRDB branch to obtain the semantic information extracted by the SCB branch, which enables the HRDB branch to focus on the detail prediction in transition region.\n"
]
},
{
"cell_type": "markdown",
"id": "ae7e1030-9044-44f5-a0d3-5148e59d4753",
"metadata": {},
"source": [
"## 5. Attention\n",
"* Training and prediction based on the public datasets Composition-1K and Distinctions-646 should be requested by email to the author.\n",
"* PP-Matting open-source pre-trained human matting model, which can be finetune using a small amount of annotated dataset according to specific human matting scenes."
" author={Chen, Guowei and Liu, Yi and Wang, Jian and Peng, Juncai and Hao, Yuying and Chu, Lutao and Tang, Shiyu and Wu, Zewu and Chen, Zeyu and Yu, Zhiliang and others},\n",
