DeepLab V3+
| Model | Training | Evaluation | Eval scales | Original | Ours (weight conversion) |
|---|---|---|---|---|---|
| DeepLab V3+ w/ Xception65 | VOC2012 trainaug | VOC2012 val | (1.0,) | 82.36 % *1 | 82.36 % |
| DeepLab V3+ w/ Xception65 | Cityscapes train fine | Cityscapes val fine | (1.0,) | 79.12 % *1 | 79.14 % |
| DeepLab V3+ w/ Xception65 | ADE20K train | ADE20K val | (1.0,) | *2 | 42.52 % |
Scores are measured by mean Intersection over Union (mIoU).
*1: Although the official repository reports a score of multi-scale prediciton, public pretrained graph is for single-scale prediction.
So we evaluated the pretrained graph using eval_semantic_segmentation in ChainerCV.
*2: Public frozen graph trained on ADE20K in official repository doesn't accept images which sizes are bigger than 513x513, while biggest image in validation set is 1600x1600.
Although we could generate a graph the biggest image can be input, it resulted 40.13% in official evaluation code.
This demo downloads a pretrained model automatically if a pretrained model path is not given.
$ python demo.py [--dataset cityscapes|ade20k|voc] [--gpu <gpu>] [--pretrained-model <model_path>] [--min-input-size <size>] <image>.jpg
Convert frozen_inference_graph.pb distributed in official repository to *.npz. Some layers are renamed to fit ChainerCV.
Official repository is here.
$ python tf2npz.py <task: {voc, cityscapes, ade20k}> path/to/frozen_inference_graph.pb <target>.npz
The evaluation can be conducted using chainercv/examples/semantic_segmentation/eval_cityscapes.py.
- Liang-Chieh Chen et al. "Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation" ECCV, 2018.
- official repository