Pytorch_Realtime_Multi-Person_Pose_Estimation
Pytorch_Realtime_Multi-Person_Pose_Estimation copied to clipboard
last-one
What is mapIdx in test_pose.py ?
Hi,
I use my datasets which has 21 key-points ( include background ) and 19 vector. Does mapIdx's length equal number of keypoints ?
If I use datasets have different number keypoints, where should I change code in test_pose.py or train_pose.py ?
Yes. The test and train scripts should be changed. I have mentioned something in README.
Thank you!
Sorry, I have another question. Why do you use 46 which has line 127 "heat_weight = 46 * 46 * 19 / 2.0" in train_pose.py ? what is 46 ?
The shape of feature (x) and the gt (y), which will be used to calculate the loss, is N * 46 * 46 (if input size is 368, 46 = 368 / 8, the receptive field of the feature is 8 in the network), and caffe's L2 loss is sum((x - y)^2) / batch_size, which is N in here. But pytorch's L2 loss is sum((x - y)^2) / the number of elements, which is N * 46 * 46 in here.
So, for comparing with origin code, the loss multiplies by 46*46. And the value varies from your input size and your network structure. If you don't change the input size and network structure, you can use 46.
Thank you
Is Input size height and weight of original image or resized image ?
Hi You said that mapIdx's length equals number of keypoints, but there is lin169 'candA = all_peaks[limbSeq[k][0]-1]' in test_pose.py. So, I think that mapIdx's length equals number of limbs.
My number of keypoints is 21 and number of limbs is 19.
Yes. The mapIdx's length equals to the number of limbs. "You said that mapIdx's length equals number of keypoints", could you tell me where I said? And I will fix it.
Thank you! I'm sorry, I misunderstood.
And, when a part isn't seen because of the oclusion, peaks is none (peaks = []). So, all_peaks's number is less than usual. What should I change part in test_pose.py ?
@tugumi911 Hi, may I ask you have you figured out the meaning of mapidx?
