demo-instructions.txt

Instructions.

This project can be broken down into three sections. Training SSD, Frame-Level Detections and Building Tubes.
Simply the SSD is trained using the training set in the dataset, Frame-Level Detections are created from the test set in the dataset and finally Tubes are build using the frame-level detections

All additional materials are available on google drive for downloading.
https://drive.google.com/drive/u/0/folders/1iF34aHNm_ygFUvrdTLq-NWDwA7PFkhBu
This includes Dataset (.jgp and labels), Pretrained weights and Detection results.

____________________________________________________________________________________
____________________________________________________________________________________
DEMO
A demo is included in the project within dataset/micro-demo-test-Set (10 frames from test video 1.1.8) this includes 11 jpg, labels and detections results.

…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…....…..…..…....…..…..…..
Frame-Level Detections (Requires Python 3.0, Pip3, a GPU, and pre-trained weights)
Download the weights: https://drive.google.com/drive/folders/1IJgcie6-aCGaZ4cktbFjVLf_iaj91pDH

SETUP Project
To install all dependancies with pip3 run:
pip3 install --user pipenv
cd realtime-action-detection/
pipenv install requests
To activate this project's virtualenv, `run pipenv shell.`

Save the weights within the datasets cache. Example: desktop/realtime-action-detection/dataset/micro-demo-test-Set/cache/CONV-SSD-oku19-rgb-bs-32-vgg16-lr-00070
Weights need to be saved in a folder like this “CONV-SSD-oku19-rgb-bs-32-vgg16-lr-00070” this means the learning rate was “7e-4” and batch-size was “32”

To evaluate the frames use this command:
CUDA_VISIBLE_DEVICES=0 python3 test-oku19.py --data_root=/Users/kcl2/Desktop/realtime-action-detection/dataset/micro-demo-test-Set --save_root=/Users/kcl2/Desktop/realtime-action-detection/dataset/micro-demo-test-Set  --input_type=rgb --eval_iter=15000 --learning-rate=7e-4 --batch_size=32

To check GPUs use “nvidia-smi”

…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..…..
Building Tubes

Open online-tubes\I01onlineTubes.m in MATLAB.

This requires data_root, save_root and iteration_nums to be declared.
If the project has been saved on Desktop this would be:
data_root = '~/Desktop/realtime-action-detection/dataset/micro-demo-test-Set';
save_root = '~/Desktop/realtime-action-detection/dataset/micro-demo-test-Set';
iteration_nums = [15000];
…
Results will be saved in save_root/results.mat. and action-path and action-tubes are also stored under save_root\* folders.
____________________________________________________________________________________
____________________________________________________________________________________

NOTES:
realtime-action-detection/dataset/[dataset-name]/splitfile contains txt files of declare what to evaluate and train on.  To change the files name to training/eval (trainval or testival)  updated (line 124) in data/oku19.py