Add SOLOv2 model

Anime_Segmentation.ipynb

@@ -4,6 +4,7 @@
   "metadata": {
     "colab": {
       "name": "Anime_Segmentation.ipynb",
+      "private_outputs": true,
       "provenance": []
     },
     "kernelspec": {
@@ -47,10 +48,12 @@
         "id": "7fpWsk8K7k3A"
       },
       "source": [
-        "# install detectron2: (Colab has CUDA 10.1 + torch 1.6)\n",
-        "# See https://detectron2.readthedocs.io/tutorials/install.html for instructions\n",
-        "assert torch.__version__.startswith(\"1.6\")\n",
-        "!pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.6/index.html\n",
+        "# install detectron2: (Colab has CUDA 10.1 + torch 1.7)\n",
+        "# tested on detectron2 v0.2-0.3\n",
+        "# See https://detectron2.readthedocs.io/tutorials/install.html for instructions if you \n",
+        "# want to switch to an older version\n",
+        "assert torch.__version__.startswith(\"1.7\")\n",
+        "!pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.7/index.html\n",
         "clear_output()"
       ],
       "execution_count": null,
@@ -78,9 +81,23 @@
         "id": "zanbjXlI79PJ"
       },
       "source": [
-        "# download model weights and config \n",
-        "gdd.download_file_from_google_drive(file_id='1-AmeAiTrtaPcNLUnRAovpveMkSxxpkKQ', dest_path='./model.pth')\n",
-        "!wget https://github.com/zymk9/Yet-Another-Anime-Segmenter/raw/main/config.yaml"
+        "# download model weights and config\n",
+        "# CondInst \n",
+        "gdd.download_file_from_google_drive(file_id='1-AmeAiTrtaPcNLUnRAovpveMkSxxpkKQ', dest_path='./CondInst.pth')\n",
+        "!wget https://github.com/zymk9/Yet-Another-Anime-Segmenter/raw/main/configs/CondInst.yaml"
+      ],
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "LNzWghZ6ziNg"
+      },
+      "source": [
+        "# SOLOv2\r\n",
+        "gdd.download_file_from_google_drive(file_id='1-wFdQ4jwSTeJ7wGD3YKNJdcpSS5Ho8c9', dest_path='./SOLOv2.pth')\r\n",
+        "!wget https://github.com/zymk9/Yet-Another-Anime-Segmenter/raw/main/configs/SOLOv2.yaml"
       ],
       "execution_count": null,
       "outputs": []
@@ -92,10 +109,10 @@
       },
       "source": [
         "# do inference\n",
-        "!python demo/demo.py --input ./input.jpg --config-file ./config.yaml --output /content/output.jpg --opts MODEL.WEIGHTS ./model.pth"
+        "!python demo/demo.py --input ./input.jpg --config-file ./SOLOv2.yaml --output /content/output.jpg --opts MODEL.WEIGHTS ./SOLOv2.pth"
       ],
       "execution_count": null,
       "outputs": []
     }
   ]
-}
+}

README.md

@@ -1,18 +1,22 @@
 # Yet-Another-Anime-Segmenter
 
 ![sample1](./sample/64535234_p0_seg.jpg)
-Instance segmentation for anime characters based on [CondInst](https://arxiv.org/abs/2003.05664), using the implementation from [AdelaiDet](https://github.com/aim-uofa/AdelaiDet) and [detectron2](https://github.com/facebookresearch/detectron2).
+Instance segmentation for anime characters based on [CondInst](https://arxiv.org/abs/2003.05664) and [SOLOv2](https://arxiv.org/abs/2003.10152), using the implementation from [AdelaiDet](https://github.com/aim-uofa/AdelaiDet) and [detectron2](https://github.com/facebookresearch/detectron2).
 
 Many thanks to [AniSeg](https://github.com/jerryli27/AniSeg) created by jerryli27, as part of the dataset originates from the segmentation data provided in this [repo](https://github.com/jerryli27/AniSeg#about-the-models). The rest of the dataset is retrieved from [Pixiv](https://www.pixiv.net/) and manually annotated.
 
+A model based on SOLOv2 had been added, which generally outperforms previous CondInst model. Evaluation results may be added soon.
+
 Newer version of the model is still under development, stay tuned if you are interested.
 
 ## Usage
 ### Installation
-Both AdelaiDet and detectron2 are required. Please refer to the official guide from [AdelaiDet](https://github.com/aim-uofa/AdelaiDet#installation) and [detectron2](https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md). A [Colab tutorial](https://github.com/zymk9/Yet-Another-Anime-Segmenter/blob/main/Anime_Segmentation.ipynb) is provided.
+Both AdelaiDet and detectron2 are required. Please refer to the official guide from [AdelaiDet](https://github.com/aim-uofa/AdelaiDet#installation) and [detectron2](https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md). A [Colab tutorial](https://colab.research.google.com/drive/1d49vI0lS2MKHx-j5VVnzPUhpRrNjueHI?usp=sharing) is provided.
 
 ### Inference
-1. Download the [pretrained model](https://drive.google.com/file/d/1-AmeAiTrtaPcNLUnRAovpveMkSxxpkKQ/view?usp=sharing) and the [config file](https://github.com/zymk9/Yet-Another-Anime-Segmenter/blob/main/config.yaml).
+1. Download the pretrained model and use the corresponding config file.
+   * CondInst: [model weights](https://drive.google.com/file/d/1-AmeAiTrtaPcNLUnRAovpveMkSxxpkKQ/view?usp=sharing), [config](./configs/CondInst.yaml).
+   * SOLOv2: [model weights](https://drive.google.com/file/d/1-wFdQ4jwSTeJ7wGD3YKNJdcpSS5Ho8c9/view?usp=sharing), [config](./configs/SOLOv2.yaml).
 
 2. Run inference with
    ```bash
@@ -22,9 +26,9 @@ Both AdelaiDet and detectron2 are required. Please refer to the official guide f
         --opts MODEL.WEIGHTS path/to/pretrained/model
    ```
 ## Training and Results
-Training using transfer learning from a [pretrained model](https://github.com/aim-uofa/AdelaiDet#coco-instance-segmentation-baselines-with-condinst) on COCO Instance Segmentation. Parameters can be found in the [config file](https://github.com/zymk9/Yet-Another-Anime-Segmenter/blob/main/config.yaml).
+Training using transfer learning from [pretrained models](https://github.com/aim-uofa/AdelaiDet/tree/master/configs/SOLOv2) on COCO Instance Segmentation. Parameters can be found in the [config file](./configs/SOLOv2.yaml).
 
-Dataset is augmented by placing segmentations on pure backgrounds.
+Dataset is augmented by placing segmentations on pure backgrounds. Models are trained with multi-scale augmentation.
 
 ![sample2](./sample/52206792_p0_seg.jpg)
 ![sample3](./sample/64113607_p0_seg.jpg)