ææ°ã®ç©ä½æ¤åºæ å ±ï¼2022/1/1追è¨ï¼
ããã®è¨äºãã3å¹´è¿ãåã®è¨äºã¨ãªããææ°ã®æ å ±ããæ¯ã¹ãã¨æ å ±ãå¤ããªã£ã¦ãã¾ãã¾ãããææ°ã®ç¶æ³ã«é¢ãã¦ã¯ä»¥ä¸è¨äºãã¨ã¦ãåèã«ãªãã¾ãã
ã以ä¸ã®è¨äºããéå»ã®æµããªã©ã¯åèã«ãªãã¾ãããã¾ã 使ããé¨åãå¤ãããã¨æãã¾ãã®ã§ããããããã°åèã«ãã¦ã¿ã¦ãã ããã
ç©ä½æ¤åºããã£ã¦ã¿ãåã«æ¤åºã¨èªèã®éã
ãããã¾ã§ããã£ã¼ãã©ã¼ãã³ã°ã使ã£ã¦ç»åã®èªèãä½åº¦ããã£ã¦ãã¾ããï¼ä»¥ä¸åç §ï¼ã
ãç»åèªèã®æ¬¡ã¯ãç©ä½æ¤åºã«æãåºãã¦è¦ãããªã¨ãããã¨ã§ããã£ã¼ãã©ã¼ãã³ã°ã使ã£ãç©ä½æ¤åºã«é¢ãã¦èª¿ã¹ã¦è©¦ãã¦ã¿ããã¨ã«ãã¾ããã
ããããããç©ä½æ¤åºã£ã¦ä½ã§ãèªèã¨ä½ãéãã®ãã¨ããã¨ãããããèªèã¨ããè¨ãã¨çµæ§åºãæå³ã«ãªã£ã¦ãã¾ã£ã¦ãç»åã®ãã®ãã®ãä½ããå¤å¥ããã®ã¯ç»ååé¡ã¨ããã®ãæ£ãããã§ããã¤ã¾ããç§ããã£ãä¸è¨ã®ä¾ã¯åºæ¬çã«ã¯ç»ååé¡ã¨ãªãã¾ããããã«å¯¾ãã¦ãç©ä½æ¤åºã¯å¯¾è±¡ãã©ãã«ããããæ¢ãè¡çºã«ãªãã¾ãããã ãæ¤åºã®å ´åã¯ãæ¤åºãããã®ãä½ããåé¡ããã¨ããã¾ã§å«ãå ´åãããããã§ããã
ãä¾ãã°ä»¥ä¸ã®åçã§ããã³ã¹ã¿ããã ï¼ã¨ããã®ãèªèã§
ãããªã¼ç´ æ
ã以ä¸ã®ããã«ããã³ã¹ã¿ããã®ä½ç½®ã示ãã®ãç©ä½æ¤åºã¨ããã¤ã¡ã¼ã¸ã§ãï¼å¤åï¼ã
ãã¡ãªã¿ã«ãä¸è¨åçã¯ãã¦ã³ãã£ã³ã°ããã¯ã¹ï¼bounding boxï¼ã¨å¼ã°ããæ ã§ç©ä½ã®ä½ç½®ã示ãã¦ãã¾ãããããããã¯ã»ã«åä½ã§é åãæå®ã§ããã¨ãã»ãã³ãã£ãã¯ã»ã°ã¡ã³ãã¼ã·ã§ã³ã¨å¼ã°ããæè¡ã«ãªãããã§ããé£ãããã§ããã
ãä¸çªæåãªç©ä½æ¤åºã¯ãé¡æ¤åºã§ãããããHarrlikeï¼ãã¼ã«ã©ã¤ã¯ï¼ã¨ããé¡ãæ¤åºããã®ã«æå©ãªç¹å¾´éãç¨ãã¦æ¤åºããã®ã代表çãªæ¹æ³ã§ããé¡æ¤åºã«é¢ãã¦è©³ç´°ã¯ã以ä¸è¨äºåç §ãã ããã
ãã¨ãããã¾ã§ã ãã ãæ¸ãã¦ãã¾ããããããå°ã詳ããç¥ãããæ¹ã¯ä»¥ä¸ã®è¨äºãè¦ãã®ãåãããããã§ããããã¾ã§æ¸ãã¦ããä¸è¨è¨äºã®åå¨æ°ã¥ãã¾ããã http://ai-4-u.com/tech/what-is-semantic-segmentation
ãèªèãåé¡ãæ¤åºãã»ã°ã¡ã³ãã¼ã·ã§ã³ãªã©ã¯ãçµæ§å®ç¾©ããã£ã¡ãã«ãã¦ä½¿ã£ã¡ããå ´åãå¤ãã®ã§ãæ£ãã使ãåããå¿ããããã§ããã
ãã£ã¼ãã©ã¼ãã³ã°ã使ã£ãç©ä½æ¤åº
ãé¡æ¤åºã®ãããªãç¹æ®ãªç¹å¾´éã使ã£ãæ¤åºã¯ãæ§è½ãé度ã§æå©ãªã®ã§ãããé¡ä»¥å¤ã«ã¯ä½¿ããªãã¨ããæ¬ ç¹ãããã¾ãããããªæ¬ ç¹ãã«ãã¼ããããã«æè¿æ³¨ç®ãæµ´ã³ã¦ããæè¡ããã£ã¼ãã©ã¼ãã³ã°ã使ã£ãç©ä½æ¤åºã§ãã
ããã£ã¼ãã©ã¼ãã³ã°ã«ããç©ä½æ¤åºã®ç 究ã®æµãã«é¢ãã¦ã¯ã以ä¸ã¹ã©ã¤ããåãããããã£ãã§ãã
ãä¸è¨ã®ã¹ã©ã¤ããä»ã®è³æãªã©ã¿ã¦ãç©ä½æ¤åºã®æè¡ã®çºå±ãèªåãªãã®ç解ãç°¡åã«ã¾ã¨ããå³ã以ä¸ã§ãã
ã大ãã3ã¤ã®æ®µéã«åããã¦ãã¦ããé åæ¢ç´¢ããã¦ãæ¤åºããç©ä½ã«åããã¦ãç¹å¾´éæ½åºããã¦ãé©åãªãæ©æ¢°å¦ç¿ææ³ããé¸æãã¦ãã¨ç©ä½æ¤åºã3ã¤ã®ã¢ã«ã´ãªãºã ã«åãã¦å®ç¾ãã¦ããã®ãåæã®æ®µéã§ããç¹å¾´éããåºæ¬ã¯æ¤åºãã対象ã«åãããå°ç¨è¨è¨ã«ãªããããç¹å®ã®å¯¾è±¡ï¼é¡ã¨ãï¼ããæ¤åºã§ããªãã®ãåºæ¬ã§ãã
ã第2段éã§ã¯ããã£ã¼ãã©ã¼ãã³ã°ã使ããã¨ã§ç¹å¾´éæ½åºãèªåã§å®ç¾ã§ããããã«ãªãããã¼ã¿ã«å¿ãã¦è²ã ãªç©ä½ãåé¡ã§ãããããã¬ã¤ã¯ã¹ã«ã¼ãèµ·ããã¾ããããã£ããï¼
ããã ãããã§ããé åæ¢ç´¢ã®ã¨ãããé ããããã¨ããåé¡ããã£ãã®ã§ãããããããããã£ã¼ãã©ã¼ãã³ã°ã§ãã£ã¡ããã°è¯ããããï¼ãã¨ããã®ã第3段éãããä»ã¯ããã£ã¼ãã©ã¼ãã³ã°ã§æåããæå¾ã¾ã§åºæ¥ã¡ããããã«ãªã£ãã®ã§ããããã¦ããã®é²æ©ããã®æ°å¹´ã¨ããããæãããä¸çã§ãã
ããã®ããã«ããã¼ã¿å ¥åãããå¾ã¯ãã£ã¼ãã©ã¼ãã³ã°ã ãã§æå¾ï¼åºåçµæï¼ã¾ã§ãã£ã¦ãã¾ããã¨ãã端ãã端ã¾ã§ã¨ããæå³ã§End-to-Endã¨è¨ãããã§ãã
ããã£ã¼ãã©ã¼ãã³ã°ã§End-to-Endã§ç©ä½æ¤åºã§ããæåãªã½ããã¦ã§ã¢ã¨ãã¦ã¯ãFaster RNNããYoloããSSDãçãããã¾ãã
ããFaster RNNãã«é¢ãã¦ã¯ã以ä¸è¨äºãããããããã£ãã§ããååã®éããFast RNNããæ©ãã®ã§Fasterã®ããã§ãã
ãããã¦ããã®Faster RNNããæ´ã«æ©ãã¨è¡æãä¸ããã®ããYoloãã§ããYoloã«é¢ãã¦ã¯ã以ååãåãããDarknetãã¨ãããã£ã¼ãã©ã¼ãã³ã°ã®ãã¬ã¼ã ã¯ã¼ã¯ã§ç¨ãããã¦ããæè¡ã§ããã¡ãã£ã¨ã ãæ¤åºã試ãã¦ããã®ã§ã使ãæ¹ãæ¦ç¥ãªã©ã¯ä»¥ä¸åç §ä¸ããã
ãYolo v2ã¯ã以ä¸ã®ãããªKeras + TensorFlowã§ã®å®è£ ãçµæ§ããã®ã§ãç°¡åã«è©¦ããããªé°å²æ°ãããã¾ããã
GitHub - allanzelener/YAD2K: YAD2K: Yet Another Darknet 2 Keras
ãããã¦ããã®Yoloã«å¹æµããé度ã§ç²¾åº¦ãé«ãï¼ãã©ã´ã³ãã¼ã«ä¸¦ã¿ã®ã¤ã³ãã¬ï¼ã¨è¨ããã¦ããã®ãããSSDï¼Single Shot MultiBox Detectorï¼ããããã§ãã
SSDã§ã®ç©ä½æ¤åºã試ãã¦ã¿ã
ãã¨ããããã§ãä»åã¯ä»ã¾ã§è©¦ãããã¨ã®ãªããSSDãã試ãã¦ã¿ããã¨ã«ãã¾ããã以ä¸ã®è¨äºãåç»ãæ¤åºããæ¹æ³ãç´¹ä»ãã¦ä¸ãã£ã¦ããã®ã§ãåºæ¬çã«ã¯ä»¥ä¸ã®ãµã¤ããåèã«ãå°ãã ã使ãåæè¯ããããã®ã使ãã¾ããã
ãããã°ã©ã ã®ãã¼ã¹ã¯ã以ä¸ã®TensorFlowå®è£ ã®SSDãforkãã¦ãã¾ãã
äºåæºå
ãPythonã§ç°å¢ãã»ããã¢ãããã¾ãããã詳細ã¯ä»¥ä¸è¨äºãªã©åç §ä¸ããã
ã代表çãªã©ã¤ãã©ãªãªã©ã®ãã¼ã¸ã§ã³æ å ±ã¯ä»¥ä¸ã§ãã
- Anaconda3==4.4.0
- tensorflow==1.3.0
- opencv-python==3.3.0.10
ãç°å¢ãã»ããã¢ããã§ãããã以ä¸ã³ãã³ãã§ãã¡ã¤ã«ã®ãã¦ã³ãã¼ãã¨ãå¦ç¿ã¢ãã«ã®è§£åãå®æ½ãã¾ãã
$ git clone https://github.com/karaage0703/SSD-Tensorflow $ cd SSD-Tensorflow/checkpoints $ unzip ssd_300_vgg.ckpt.zip $ cd ..
SSDå®è¡
ãMacBookã®å ´åã¯ã以ä¸ã³ãã³ããå®è¡ããã¨MacBookå èµã«ã¡ã©ã®ç»åã§ãªã¢ã«ã¿ã¤ã ã«æ¤åºãèµ°ãã¾ããCore i5ã®MacBook Proã§2ç§ç¨åº¦ã§æ¤åºã§ãã¾ãã
$ python ssd_camera.py
ãoutput.avi
ã¨ãããã¡ã¤ã«ã«ãæ¤åºçµæãæ¸ãåºããã¾ãã
ãã«ã¡ã©ã®ä»£ããã«ãåç»ã使ç¨ãããã¨ãå¯è½ã§ããä¾ãã°ã好ããªã«ã¡ã©ã§åç»ãæ®ã£ã¦sample.mov
ã¨ãããã¡ã¤ã«åã§ä¿åããå ´åã¯ã以ä¸ã®ããã«ç¬¬ä¸å¼æ°ã«åç»ãã¡ã¤ã«åãå
¥ãã¦ã³ãã³ãã®å®è¡ããã¨åç»ããæ¤åºã§ãã¾ããåç»ã¯ã人ãåã£ã¦ããã¨çµæãåãããããã§ãã
$ python ssd_video.py sample.mov
ãçµæã¯ãoutput.avi
ã¨ãããã¡ã¤ã«ã«ä¿åããã¾ãããããæ¢ã«output.avi
ã¨ãããã¡ã¤ã«ãããå ´åã¯ãäºãæ¶ããå¥ã®ååã«ãã¦ããã¾ãããã
ãä»åã¯ããã¤ãã®ããã«ããªã¼ç´ æã®ãã³ã¹ã¿ï¼id:lonestartxï¼ããã®ãã¹ãªããã¹ãã£ã³åçãä½ãããã«æ®å½±ããåç»ã使ç¨ãã¾ãããããªã¼ç´ æã®åå©ç¨ã§ããã¨ã£ã¦ãã¨ã³ã§ããã
ãSSDã§æ¤åºããçµæã¯ä»¥ä¸ã§ãããã³ã¹ã¿ããã®åãã«Personï¼äººéï¼ã¨ãããã¦ã³ãã£ã³ã°ããã¯ã¹ãå®å ¨ã«è¿½å¾ãã¦ãã¾ãããã³ã¹ã¿ãããå®å ¨ã«äººéã¨ãã¦æ¤åºããã¦ãã¾ãããã³ã¹ã¿ããã¯äººéã§ããï¼
ã¾ã¨ã
ããã£ã¼ãã©ã¼ãã³ã°ã«ããç©ä½æ¤åºã«é¢ãã¦ãç°¡åã«ã¾ã¨ãã¦ãSSDãã«é¢ãã¦å®è·µãã¦ã¿ã¾ããã
ãããã¦ããã®è¨äºãæ¸ãã¦å ¬éããã¾ã§ã®éï¼2é±éãããï¼ã«ãYoloã®æ°ãããã¼ã¸ã§ã³ã§ãããYolo V3ããªãã¦ã¤ãããã®ãåºã¦ãã¦ãã¾ã£ãããSSDãæ¹è¯ãã¦æ¼«ç»ã®ç»åããç©ä½æ¤åºã§ããããã«ããç 究çºè¡¨ãåºã¦ãããã¨ãç¸å¤ããããã®åéã®çºå±é度ã¯åãã§ããã¤ãã¦ããã®ã大å¤ã§ãã
ãã¨ããããã§ãYolo V3ã®ã¤ã«ãããã¢åç»ãè¦ãªããä»æ¥ã¯ãå¥ããããã¨æãã¾ãããããªãããããªãããããªãã
ã追è¨ï¼ããã«ææ°ã®ç©ä½æ¤åºã®ç 究ã«é¢ãã¦ã¯ãéå»ã®æµããå«ãã¦ä»¥ä¸ã®ã¹ã©ã¤ããåãããããã£ãã§ãã
ç©ä½æ¤åºã®å¿ç¨ä¾
ãTensorFlowã§ãæ軽ã«æ§ã ãªã¢ãã«ã使ã£ãç©ä½æ¤åºã試ãããObject Detection Toolsãã¨ãããã¼ã«ãä½æãã¾ããã以ä¸åç §ä¸ããã
ãç©ä½æ¤åºãå¿ç¨ãã¦ãèªåã§ã¢ã¶ã¤ã¯ããããããã£ã¼ãã¢ã¶ã¤ã¯ããä½ã£ã¦ã¿ã¾ããã
åèãªã³ã¯
SSD: Single Shot MultiBox Detector · Issue #659 · arXivTimes/arXivTimes · GitHub
YOLOv2(Keras / TensorFlow)でディープラーニングによる画像の物体検出を行う - Qiita
Deep Learningを使用した物体検出方法の紹介 | Sosogu LLC.
リアルタイム物体検出向けニューラルネット、SSD(Single Shot Multi Detector)及びその派生モデルの解説 - verilog書く人
What’s new in YOLO v3?. A review of the YOLO v3 object… | by Ayoosh Kathuria | Towards Data Science
ChainerCVとLight-Head R-CNNで『カメラ・動画対応!物体検出ソフト』を作る|はやぶさの技術ノート
CNNを用いた物体検出アルゴリズムYOLOv3に迫る! - Kysmo’s Tech Blog
Going deep into object detection. With recent advancements in deep… | by Lars Hulstaert | Towards Data Science
Semantic segmentation 振り返り - Speaker Deck
Introduction to Object Detection. From Image Classification to Object… | by Mahendran Venkatachalam | Towards Data Science
【物体検出】mAP ( mean Average Precision ) の算出方法 - Qiita
ディープラーニングの物体検出のサーベイ論文まとめ - からログ
Object Detection — STAT 157, Spring 19 documentation
é¢é£è¨äº
å¤æ´å±¥æ´
- 2022/01/01 ææ°ã®æ å ±ã«é¢ãããªã³ã¯è¿½è¨
- 2020/06/14 åèãªã³ã¯è¿½å
- 2020/06/13 åèãªã³ã¯è¿½å
- 2020/05/19 åèãªã³ã¯è¿½å
- 2020/04/04 é¢é£ãªã³ã¯ã追å
- 2019/06/29 å¿ç¨ä¾ã«é¢ãã¦è¿½è¨
- 2019/04/12 åèãªã³ã¯ãã¾ã¨ãã«ææ°ã®ç 究ã«é¢ãã¦è¿½è¨
- 2018/12/13 åèãªã³ã¯è¿½å