SCKD : Self-Cross Similarity Knowledge Distillation for Light Camera Pose Regressor

The similarity based knowledge distillation mehod to compress the 6-dof Pose Regressor model so that it can show high performance on 5W low-power environment

Architecture

Requirements

• LINUX
• Colab
  • Python 3
  • CPU : Intel Xeon CPU 2.3 GHHz(Dual-Core)
  • GPU : Nvidia Tesla T4
  • GPU memory : 8GB

Installation

git clone https://github.com/e-LENS/SCKD.git
cd SCKD
pip install -r requirements.txt

Dataset

• Outdoor dataset

  Download King's College dataset under datasets folder

• Indoor dataset

  Download 7 Scenes dataset under datasets folder

datasets/KingsCollege
datasets/Chess
datasets/Fire
datasets/Heads
datasets/Office
datasets/Pumpkin
datasets/RedKitchen
datasets/Stairs

• Compute the mean image for each dataset

python util/compute_image_mean.py --dataroot datasets/[Dataset_name] --height 256 --width 455 --save_resized_imgs

Preprocess 6DoF

In 7 Scenes dataset, Position & Orientation is given in 4x4 matrix format.
PoseNet requires 7 dimension vector of position (X,Y,Z) & Orientation quaternion (W,P,Q,R)

Use posenet-pytorch/7scenes_preprocessing.py to convert 4x4 matrix to 7 dim vector.

SCKD

Train & Test the PoseNet model on each dataset

Pretrained model

'resnet18im': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
'resnet34im': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
'resnet50im': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
'resnet101im': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
'resnet152im': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',

Train Teacher Model

!python train.py --model [resnet18 | resnet34 | resnet50 | resnet101] --dataroot [DATAROOT] --name [model]/[Dataset]/[beta500_ex] --beta 500 --gpu 0 --niter 500 --batchSize32 --lr 0.001

Test Teacher Model

!python test.py --model [resnet18 | resnet34 | resnet50 | resnet101] --dataroot [DATAROOT] --name [model]/[Dataset]/[beta500_ex] --gpu 0

Train Student Model

!python KD_train.py --model [resnet18 | resnet34 | resnet50 | resnet101] --dataroot [DATAROOT] --name [student model]/[Dataset]/[beta500_bt_lr_m#1_m#2_scaling] --beta 500 --gpu 0 --niter 500 --T_model [ resnet34 | resnet50 | resnet101]  --T_path [TeacherModel_Path] --save_epoch_freq 5 --SCmodule [ 0 | 1 | 2 | 3 | 4 | 5 ] --hintmodule [ 0 | 1 | 2 | 3 | 4 | 5 ] [--SCKL]

• SCmodule 과 hintmodule option layer는 list type 으로 선택가능

Test Student Model

!python KD_test.py --model [resnet18 | resnet34 | resnet50 | resnet101] --dataroot [DATAROOT] --name [student model]/[Dataset]/[beta500_bt_lr_m#1_m#2_scaling] --beta 500 --gpu 0 

• Train result is saved here ; ./checkpoints/[name]
• Test result is saved here ; ./results/[name]

Result

** ResNet Model Size **

Backbone Models	ResNet50(Teacher)	ResNet34(Teacher)	ResNet18(Student)
Size(MB)	102.45	87.27	46
Number of Parameter	25,613,383	21,817,159	11,708,999

** Inference time at Jetson Nano 2GB developer kit **

Backbone Models	ResNet50(Teacher)	ResNet34(Teacher)	ResNet18(Student)
Median Inference Time (clock)	125	90	50

** Comparision with other KD method for Regression problems **

Model	Shop Facade	Stairs
Vanilla KD	8.25m / 9.17°	0.39m / 13.08°
M.U's KD	1.45m / 7.58°	0.39m / 15.27°
Self-Similarity KD	0.94m / 6.56°	0.37m / 13.63°
Ours	0.93m / 5.96°	0.33m / 13.02°

---

Reference

• https://github.com/hazirbas/poselstm-pytorch
• https://github.com/HobbitLong/RepDistiller