gzerveas / gaila_object_detection

A modification of CenterNet for a VR graphics dataset (GAILA)

Object Localization for Virtual Reality Environments

This repo contains the code for object detection in VR environments. The method is based on CenterNet and is applied on GAILA dataset.

CenterNet paper: Objects as Points,
Xingyi Zhou, Dequan Wang, Philipp Krähenbühl,

Installation

Please refer to INSTALL.md for installation instructions.

Use

To build the dataset and convert the raw data to proper format run the following:

python main.py gaila_ctdet --exp_id TEST_gaila_simplenet  --frames_dir ~/data/GAILA/images_10hz/ --bounds_dir ~/data/GAILA/bounds/ --save_annotations ~/scratch --arch simple --batch_size 32 --master_batch -1 --lr 1.25e-4 --gpus 0 --num_workers 4  --frames_per_task 20 --num_epochs 2 --debug 4

For training the model run:

python  main.py gaila_ctdet --exp_id TRAIN_gaila_resdcn18_fpt400_ep10  --load_annotations ../data/ --arch resdcn_18 --batch_size 128 --master_batch -1 --lr 1.25e-4 --gpus 0 --num_workers 8 --frames_per_task 400 --num_epochs 100 --resume

For testing the model run:

python gaila_eval.py gaila_ctdet --exp_id TEST_gaila_resdcn18_fpt400_ep10 --vis_thresh 0.4 --eval_vis_output ../exp/output_dump/ --load_annotations ../data --batch_size 32 --num_workers 8 --master_batch -1 --arch resdcn_18 --load_model ../exp/gaila_ctdet/TRAIN_gaila_resdcn18_fpt400_ep10/model_last.pth