Experiment Presentation YOLO 9000 Better Faster Stronger RFCN

Experiment Presentation YOLO 9000: Better, Faster, Stronger & R-FCN: Object Detection via Regionbased Fully Convolutional Networks -Yingzhou

Experiment Presentation YOLO 9000: Better, Faster, Stronger

Characteristics - Can detect more than just 200 classes; it predicts detections for more than 9000 different object categories - Use a custom network based on the Googlenet architecture, use Word. Tree to combine data from various sources and propose a joint optimization technique to train simultaneously on Image. Net and COCO - Outperforming methods like Faster R- CNN with Res. Net and SSD while still running significantly faster

Dataset-paper • Training - train YOLO 9000 simultaneously on the COCO detection dataset and the Image. Net classification dataset • Testing - PASCAL VOC datasets (78. 6 m. AP on the 2007 set) • Speed - uses a custom network based on the Googlenet architecture. This network is faster than VGG-16, only using 8. 52 billion operations for a forward pass.

Dataset-test • Category: 50 pictures of different categories, including human being, animal, food, etc • Image pixel: various, 500 × 375, 450 × 321, 1600 × 1067 …

Environment • • • Processor: Intel Core i 7 -4910 QM Memory: 32 G LDDR 3 2133 MHZ Graphics: NVIDIA GTX 980 M System: Linux Software: Matlab

Results • Detection: can detect most objects and report m. AP (Mean Average Precision) for detected objects, but there are still many false recognition • Time: mean time: 6. 221 s (CPU) maximum: 6. 399 s (CPU) minimum: 6. 086 s (CPU)

Example-strength threshold = 0. 25, running time = 6. 358 s (CPU)

Example-strength threshold = 0. 5, running time = 6. 185 (CPU)

Example-weakness threshold = 0. 25, running time = 6. 183 s (CPU)

Example-weakness

Example-weakness Tigon threshold = 0. 25, running time = 6. 399 s (CPU)

Example-weakness threshold = 0. 5, running time = 6. 323 s (CPU)

Analysis accurate and efficient for most object detection, but there are still many false recognition for similar animals

Conclusion • Using multi-scale training YOLOv 2 model can run at varying sizes, offering an easy tradeoff between speed and accuracy.

Experiment Presentation R-FCN: Object Detection via Regionbased Fully Convolutional Networks

Model

Dataset-paper • Training - train the models on the union set of VOC 2007 trainval and VOC 2012 trainval • Testing - PASCAL VOC datasets (83. 6% m. AP on the 2007 set) with the 101 -layer Res. Net. • Speed - the running time per image is 170 ms on one Nvidia K 40 GPU, 2. 5 -20× faster than the Faster R-CNN counterpart

Dataset-test • Category: pictures of 8 different categories, including human being, horse, sheep, dog, vehicle(car & bus), cat • Image pixel: each picture 500 × 375

Environment • • • Processor: Intel Core i 7 -4910 QM Memory: 32 G LDDR 3 2133 MHZ Graphics: NVIDIA GTX 980 M System: Windows 10 x 64 Software: Matlab + Caffe

Results • Detection: can detect most objects and report m. AP (Mean Average Precision) for detected objects • Time: mean time: 0. 211 s maximum: 0. 222 s (sheep) minimum: 0. 205 s (human & dog)

Example-strength running time = 0. 206 s (GPU)

Example-strength running time = 0. 207 s (GPU)

Example-weakness running time = 0. 215 s (GPU)

Example-weakness

Analysis • accurate and efficient for most object detection, but may ignore some small and vague one • faster than traditional method, also surpass R -CNN(0. 42 s per image)

Conclusion • The algorithm is very robust for object detection, it achieves accuracy competitive with the Faster R-CNN counterpart • Outstanding in reducing running time