rl_pose_estimation's Introduction

Reinforcement Learning Multi-Person Pose Estimation

Final Project for Vision Meets Machine Learning Trying to apply Q-learning on Multi-Person Pose Estimation The goal is to use Q-learning to recognize the human action Clapping It is based on the paper CVPR'17 paper and its GitHub link.

The human action detection dataset is from NTU RGB+D dataset.

Codes

read_skeleton_file.py

Arrange skeleton data to what we need, discard some joints we don't need.

rl_qtraining.py

Apply Q-Learning on to the data, train the agent to recognize clapping.

Approaches

Use Traditional Q-Learning
Setup a game (move all skeletons) and an agent (try to do clapping)
Inputs: States (13 joints * 2 features(v, d) * 3 coordinates)
Outputs: Actions (13 * 1 * 3)
Model

    model = Sequential()
    model.add(Dense(hidden_size, input_shape=(2*13*3,), activation='relu'))
    model.add(Dense(hidden_size, activation='relu'))
    model.add(Dense(13*3))
    model.compile(loss='mse', optimizer='sgd')
    self.Q = model

Agents remembers previous moves
Give reward for each action according to the frame number and do training at the end

rewards.append([[frame], [self.env.scaling_factor * pow(self.discount,self.env.max_frame - frame)*reward]])

Gamma (discount) = 0.95
Training

    def replay(self, batch_size):
        inputs, targets, rewards= self._prep_batch(batch_size)
        # the model is trained here
        print(rewards.shape)
        loss = self.Q.train_on_batch(inputs, targets)
        return loss

Recommend Projects