gitxub / trinet

TriNet for 3DOF object pose estimation

TriNet

TriNet for 3DOF object pose estimation

Network Overview

Train

python train.py --train_data --valid_data --basenet --alpha_reg --beta --num_bins --batch_size --save_dir --ortho_loss --cls_loss --reg_loss

Test

python test.py --test_data --snapsot --num_classes --model_name

Video Test

./video.sh

Webcam Test

./webcam.sh

Dependencies:

matplotlib==3.1.3
numpy==1.18.1
opencv-python==4.1.1.26
pandas==1.0.1
pillow==7.0.0
pyparsing==2.4.6
python==3.7.6
scipy==1.4.1
seaborn==0.10.0
setuptools==45.2.0
sqlite==3.31.1
tensorboardx==2.0
tk==8.6.8
torch==1.4.0
torchvision==0.5.0
tqdm==4.43.0

Data Pre-processing

Three datasets are involved: BIWI, AFLW, 300W_LP. We describe how we obtain three directional vectors in data/computeMAE.py file

Vector Refinement

We refine three network output vectors by solving a optimization problem. See implementation details in Optimization/optimize.py

Experiments

Comparison to state-of-the-art methods on AFLW2000

Method	roll	pitch	yaw	MAE
Dlib	10.545	13.633	23.153	15.777
3DDFA	8.250	8.530	8.5400	7.393
Hopenet	5.674	6.637	6.920	6.410
Quatnet	3.920	5.615	3.973	4.503
FSA_Caps_Fusion	4.64	6.08	4.50	5.07
TriNet(mobilenetv2)	2.86	5.25	5.34	4.48
TriNet(resnet50)	2.61	4.25	4.70	3.86

Comparison to state-of-the-art methods on BIWI

Method	roll	pitch	yaw	MAE
Dlib	9.324	15.505	20.581	15.137
3DDFA	11.770	11.180	8.691	10.547
KEPLER	16.196	17.277	8.691	10.547
Hopenet	3.269	6.606	4.810	4.895
Quatnet	2.936	5.492	4.010	4.146
FSA_Caps_Fusion	2.76	4.96	4.27	4.00
TriNet(mobilenetv2)	6.94	6.75	4.02	5.90
TriNet(resnet50)	3.67	4.99	3.27	3.97

Error distribution testing on AFLW2000

Testing on AFW