StarGAN

Ref:
https://github.com/yunjey/stargan
https://arxiv.org/pdf/1711.09020.pdf

StarGAN如圖(b)與其他GAN模型相較, 受到關注的特色在於其不同Domain間的轉換可以使用同一個模型, 不需要一對一Domain的產生很多組模型, 如圖(a).

Picture originated from: https://arxiv.org/pdf/1711.09020.pdf

下圖是使用StarGAN生成的圖片, 一組有六張, 分別是

Input + Black_Hair +Golden_Hair +Brown_Hair +Gender_Change +Aged 共5個Domain

Train這個模型在FX705GE(CPU:INTEL i7-8750H, 32G RAM)上花了約39小時作Training, 總共200,000 steps, 執行速度7 sec/ 10steps ,

相對GL753VE(CPU:INTEL i7-7700HQ, 24G RAM) 90 sec/ 10steps <== 只是worker node, 未開啟NVIDIA CUDA

從下列image看來效果有好有壞, 視人頭比例, training dataset, 相片品質及背景等條件而異.

感謝同仁們(含EX-)及致中提供玉照協助!! 使用celebrity dataset作training 及部份testing

Input + Black_Hair + Golden_Hair + Brown_Hair + Gender_Change + Aged

留言

這個網誌中的熱門文章

OCR應用在電子元件上的辨識

OCR Application Example1: for SMD idenfication : Text detect by CRAFT OCR文字偵測原始照片為網路上下載，再套上OCR文字偵測顯示結果，若有侵權請告知移除彩色區域為偵測到文字的部份 Output 10 coordinates of corresponding text blocks 1. 144,196,286,194,287,259,145,261 2. 298,198,509,196,509,259,298,262 3. 148,262,286,262,286,321,148,321 4. 368,266,513,264,513,321,369,323 5. 145,331,472,333,471,395,145,393 6. 146,404,445,404,445,454,146,454 7. 146,453,512,453,512,502,146,502 8. 147,502,481,499,481,551,148,553 9. 148,550,614,550,614,600,148,600 10.513,600,714,600,714,648,513,648 After image pre-processing: OCR result1: After image pre-processing: OCR result2: Example2: for datasheet interpretation : Text detect of TI datasheet by CRAFT OCR results: ([[75, 11], [127, 11], [127, 31], [75, 31]], 'TEXAS', 0.999188403930061) ([[474, 4], [928, 4], [928, 32], [474, 32]], 'PACKAGE MATERIALS INFORMATION', 0.6743955072876302) ([[77, 29],...