Computer Vision Course Projects

This project is about tasks for Computer Vision Course in Electronic Information and Communication Department, Huazhong University of Science and Technology, and the teacher is Prof. Xinggang Wang.

Project navigation

name	usage
Proj1	perceptron for linear classification
Proj2	one layer neural network for classification
Proj3	multi-layer neural network for mnist classification
Proj4	pytorch version for mnist or cifar10 classification
Proj5	pytorch version for single object localization
Proj6	pytorch version for sematic segmentation

---

Proj1 - perceptron

---

Proj2 - neural network

---

Proj3 - mnist

file	usage
function.py	define the activate function such as ReLU, softmax and sigmoid
Net.py	define the network structure and forward/backward process
mnist.py	how to load data and label from .gz files
main.py	batch learning and figure plotting

Different activate functions will contribute to different output

Different init principle will contribute to different output

---

Proj4 - pytorch classification

file	usage
MyNet.py	generate a simply CNN model
ResNet.py	transfer learning from Res50
VGG.py	transfer learning from VGG16
Plot.py	plot the final results, such as loss and acc

note: In ./pytorch/models I only store three models, because Res50 and VGG16 models are too huge to store in Github

---

Proj5 - pytorch single localization

file	usage
Net.py	VGG16 pretrained network + customize
VGG_loss_weight.py	main file, use two loss funcs (cross entropy and Smooth L1)
VGGConV_xxx.py	using ConV instead fc
visualization.py	plot the final results
dataloader.py	load data from file as ndarray and tensor
tiny_vid.tar	dataset sampled from VID

The following figures are "weight=1:1", "weight=1:1e-1", "weight=1:2e-2", "ConV network"

---

Proj6 - pytorch sematic segmentation

file	usage
TinySeg	dataset sampled from VOC 2012
backbone_8stride.py	ResNet backbone structure
eval_seg.py	evaluation code
train_seg.py	train code
test_seg.py	plot the output imgs
pspnet.py	PSP module and network
sync_batchnorm	syncronize batchnormal in DeepLab
DeepLab.py	define the deeplab v3 network
train_seg_deeplab.py	train stage for deeplab net

PSPNet

H.Zhao, J.Shi, X.Qi, X.Wang and J.Jia. Pyramid Scene Parsing Network. CVPR2017

• mIoU is about 0.72, while the SOTA of PSPNet in VOC2012 is 0.82

output samples

DeepLab v3

L.Chen, G.Papandreou, F.Schroff, H.Adam. Rethinking Atrous Convolution for Semantic Image Segmentation. arXiv:1706.05587v3

• mIoU is about 0.74

---