README_en.md

Model Name

DenseDepth

1.Introduction

DenseDepth is form Arxiv: High Quality Monocular Depth Estimation via Transfer Learning, used for monocular depth estimation.

High Quality Monocular Depth Estimation via Transfer Learning

DenseDepth mainly extracts image features through Densenet169 encoder and bilinear interpolation up-sampling decoder which restores image depth. The main idea is to use the backbone trained on ImageNet, carrying out transfer learning. In addition, the training direction is constrained by SSIM luminosity error, image gradient smoothing error and depth estimation error during training.

On NYU depth dataset densedepth achieves SOTA at monocular depth estimation.

2.Reproduce performance

	a1	a2	a3	rel	rms	log10
origin repo	0.895	0.980	0.996	0.103	0.390	0.043
paddle reproduce	0.896	0.981	0.995	0.106	0.456	0.044

3.NYU Depth Dataset

nyu depth dataset is used to train. Link is as below: nyu_depth

The dataset is nearly 4.1G，which contains 50k training RGB images and 1308 testing RGB images with their depth groundtruths.

4.Environments required

The software dependencies are as below：

paddlepaddle-gpu>=2.0
tensorboardX

The hardware:
training needs GPU RAM>=9GB(batch_size=4) RAM>=18GB(batch_size=8)

testing can be done on CPU but is very slow.

5.Quick start

using pretrained DenseDepth to test and eval

densenet pretrained model:
Link: https://pan.baidu.com/s/1KUPnjUgpG40VSDLBHEcRIQ
code：zid4
densenet model should be put on model/

DenseDepth pretrained model:
Link:https://pan.baidu.com/s/1f1lYptz3xVMs3mJKvVgrUw
Code:as0z
DenseDepth model should be put on logs/

then do predict or eval:

# predict your own image, you should put RGBs on images/
python predict.py

# eval the model. you should put nyu_data.zip on data/
python eval.py

using densenet pretrained model to transfer learning

densenet pretrained model:
Link:https://pan.baidu.com/s/1KUPnjUgpG40VSDLBHEcRIQ
Code:zid4
put densenet on model/

put nyu_data.zip on data/

modify configs/main.cfg(recommand to use the default params)，then start training:

python train.py

6.Code structure and Introduction

structure

├─configs
├─data
├─images
├─logs
├─model
├─results
├─utils
│ eval.py
│ predict.py
│ README.md
│ README_cn.md
│ requirement.txt
│ train.py

Instructions

configs/main.cfg train,predict,eval are used on training, testing and evaluation separately.

[train] training params 
# 
epochs = 20
# 
learning_rate = 0.0001
# 
batch_size = 8

[eval] evaluation params
# 
weights_path = logs/DenseDepth_val_best.pdparams

[predict]
# 
weights_path = logs/DenseDepth_val_best.pdparams
# 
imagedir_path = images/
# 
color_map = gray

Files instructions

configs/main.cfg   # params config
data/data.py   # dataset processing
images/   # images dir for testing
logs/   # model save dir when training
model/densenet   # densenet backbone
model/model.py   # densedepth mdoel
results/   # image dir for testing results
utils/losses.py   # losses definition on training
utils/utils.py    # some utils for visualization and image loading

eval.py  # evaluation main file
predict.py   # testing main file
train.py   # training main file