search

sfzhang15 / RefineDet

Single-Shot Refinement Neural Network for Object Detection, CVPR, 2018
Other
1.44k stars 392 forks source link

Just Train Network for Person Detection #42

Closed ahkarami closed 6 years ago

ahkarami commented 6 years ago

Dear @sfzhang15, Thank you for your work. I have a general question. In the VOC and COCO datasets there are some classes (e.g., 20 or 80 classes (person, car, horse, etc.) for object detection). If one train the network just for one class (e.g., just the person class) will the accuracy (mAP) of person detection increase?

sfzhang15 commented 6 years ago

@ahkarami Hi, We are not sure whether it will increase the accuracy. But if you want to train the method for person detection, you should use the specific datasets, i.e., CityPersons, Catech and so on. Besides, you should adjust some parameters correspondingly, i.e., anchor scales, anchor ratios, learning rate, etc. We have the plan to train our RefineDet for pedestrian detection.

ahkarami commented 6 years ago

@sfzhang15, Thank you for your reply. I hope your trained model for person detection with its interesting results prepare soon. As you have excellent knowledge and experience in object detection, I want to know that, generally if we reduce the number of classes in a data set, will this increase the accuracy of detection? (for example instead of train the model on all 20 classes in VOC data set, just train it on 1 class). However, I think, as your previous answer, you are not sure about the answer of this question. Excuse me for my inconvenience, I have two more questions. 1- Regarding your above answer, you mean that the VOC & COCO data sets are not appropriate data sets for person detection (and for example CityPersons and so on, are proper)? 2- What's your opinion about Open Images data set? [its quality for person detection task] Thank you

sfzhang15 commented 6 years ago

@ahkarami Fisrtly, I want to emphasize the difference between pedestrian detection and person detection. Pedestrian detection only focuses on walking person, which are standing and has a roughly fixed aspect ratio, such as 0.41. While person detection aims at finding all kinds of person, i.e., pedestrain, riders, sitting person, unusual person and so on. If you only want to detect pedestrian, you should use those pedestrian detection datasets, such as citypersons and caltech. If you want to detect person, you can use those datasets which has the person class.

ahkarami commented 6 years ago

Dear @sfzhang15, Thank you very much for your complete answer. Your mentioned notes was really helpful for me.