关于Tensorflow Object Detection API注意事项

Lessons learned（经验教训）

Some key findings from the Google Research paper: 谷歌研究报告的一些重要发现：

• R-FCN and SSD models are faster on average but cannot beat the Faster R-CNN in accuracy if speed is not a concern.
• 一般而言，R-FCN和ssd的速度更快，但如果不考虑速度的话，其精度无法超过Faster R-CNN。
  Faster R-CNN requires at least 100 ms per image.
• Faster R-CNN每张图像至少要100毫秒。
• Use only low-resolution feature maps for detections hurts accuracy badly.
• 仅使用低分辨率特征图会严重降低准确度。
• Input image resolution impacts accuracy significantly. Reduce image size by half in width and height lowers accuracy by 15.88% on average but also reduces inference time by 27.4% on average.
• 输入图像的分辨率对准确度影响很大。将图像的宽度和高度减少一半，平均降低了15.88%的准确度，同时也平均减少了27.4%的推理时间。
• Choice of feature extractors impacts detection accuracy for Faster R-CNN and R-FCN but less reliant for SSD.
• 对于Faster R-CNN和R-FCN，特征提取器的选择会影响检测准确度，但SSD对此依赖性较小。
• Post processing includes non-max suppression (which only run on CPU) takes up the bulk of the running time for the fastest models at about 40 ms which caps speed to 25 FPS.
• 后期处理包括非极大值抑制（仅在CPU上运行）占最快模型推理时间的大部分，约为40毫秒，速度上限为25 FPS。
  If mAP is calculated with one single IoU only, use mAP@IoU=0.75.
• 如果mAP仅使用一个IoU计算，则使用mAP@IoU=0.75。
• With an Inception ResNet network as a feature extractor, the use of stride 8 instead of 16 improves the mAP by a factor of 5%, but increased running time by a factor of 63%.
• 使用Inception ResNet网络作为特征提取器，使用步长8而不是16将地图提高了5%，但将运行时间提高了63%。
  Most accurate 最准确
• The most accurate single model use Faster R-CNN using Inception ResNet with 300 proposals. It runs at 1 second per image.
• 最精确的单一模型使用Faster R-CNN，使用Inception ResNet和300个候选框。它以每张图像1秒的速度运行。
• The most accurate model is an ensemble model with multi-crop inference. It achieves state-of-the-art detection on 2016 COCO challenge in accuracy. It uses the vector of average precision to select five most different models.
• 最精确的模型是具有多截断推断的集成模型。它在2016年COCO挑战赛上实现了最先进的检测。它使用平均精度向量来选择五种最不同的模型。
• Fastest 最快的
• SSD with MobileNet provides the best accuracy tradeoff within the fastest detectors.
• 带MobileNet的SSD在最快的检测器内提供最佳的精度权衡。
   
• SSD is fast but performs worse for small objects comparing with others.SSD速度很快，但对小对象的性能比其他对象差。
• For large objects, SSD can outperform Faster R-CNN and R-FCN in accuracy with lighter and faster extractors.
• 对于大型对象，SSD在精度上优于更快的R-CNN和R-FCN，具有更轻和更快的提取器。
   

Good balance between accuracy and speed 准确度和速度之间的良好平衡

• Faster R-CNN can match the speed of R-FCN and SSD at 32mAP if we reduce the number of proposal to 50.

• 如果我们将提案数量减少到50个，更快的R-CNN可以在32mAP上与R-FCN和SSD的速度相匹配。