ndcuong91/c3d_luanvan

This repo contains source code and result for master thesis in April 2019

Hi there 👋

• 🔭 I’m a technical project manager in AI Research Lab of Samsung SDS Vietnam
• 🌱 I’m really interested in application of AI in real life! ...
• 📫 How to reach me: Facebook or Email.
• ⚡ Fun fact: I'm also a speedcuber (retired!). Let see my Youtube

c3d_luanvan's Issues

Modification and roadmap information

Modification list from original code:

• Delete all .fc6, .fc7, .prob vector in result folder
• Change result folder (to locate .fc6, .fc7, .prob vector) in output/result
• Don't delete result folder (e.g. Kinect_1_test_on_Kinect_1), add time to it
• Finetuning continuously for better experiment on params
• Add arguments for more flexible when training
• Add data_augmentor.py script
• Add params.txt in output result folder which list all the parameters used for finetuning
• Train & test set can use different data_type (e.g. train on K_1_original, test on K_1_segmented)
• Can use average features for classification or only first feature (for first 16 frames)
• Check GPU ready
• Use center crop instead of random crop in training
• Add summary result
• Modify net.cpp to run with different filter in convolution

Road map (15/12/2018 --> 30/03/2019)

• Modify code, Explore dataset, Preprocessing and Tuning hypeparameters

• Data augmentation

• Modify input size for best result

• Find the best way to pre-processing (pre 3)

• Find the most suitable hyper-parameters set

• Run code with 12 hand-gesture dataset

• Finish create dataset

• Run code with 12 hand-gesture dataset

• Make simpler network

• Change loss function

• Find the best way for training that can achieve good result on every view