Article link: https://arxiv.org/pdf/1711.11575.pdf
1. What do the relationship between the object detection?
Use appearance features, and geometric information, namely: a convolutional network and features box
2. The article says that he is the end to end object detetion, why?
nms be implemented by the network
3. Does the network need other supervisory signal?
No, as a similar deformable network, no additional information
4. What a way to use this article before the relationship between objects do after treatment there?
Depth previously studied, using constrained DPM object appearing simultaneously improving the accuracy of object recognition, there are ways used to improve the accuracy of the position and size,
Depth study of the times, lstm and spatial memory networks have been proposed, but was not able to improve stateofart model, and training complex
Based on human tasks, but need to mark people's movements, this method does not require
5. This paper presents a space right weight, after conversion, will change?
Transformation constant
6. Comparison attention basic form of Equation 1, the formula 4, who is the key, who is the value?
Basic forms: formula 4:
, n is a visible feature of the box, box of all the m feature is traversed, so that n is the feature query, m is the key feature
7. If you want to get the relationship between the size of two box which formula used?
Thesis Formula 3
