darker66678/nycu_vrdl_homework3

This is Homework 3 of VRDL about Instance segmentation for necleus dataset

NYCU_VRDL_Homework3

This is Homework 3 of VRDL about Instance segmentation for necleus dataset

| cocoapi cocoAPI is a very powerful tool in Segmentation and Object detection

| mmdetection This is a reference project from mmdetection

| --- model.sh Command of model

| --- sub.sh Command of zipping answer.json

| --- train_config Pretrain model and config file are put in here

| --- tools

| ------ train.py Main training program

| ------ test.py Main inference program

| nucleus Training and testing datasets should be put in here

| coco_trans.py Program of Converting mask img

| data_collation.py Shell script for data collation

| install.sh Shell script for installing

Requirements

Using python 3.7 in Anaconda

sh install.sh Install all the dependencies

Data preprocessing

Download Nucleus dataset and put in nucleus folder

Run python data_collation.py for Collation all the data

Run python coco_trans.py for Converting mask img into json file of coco

Training

you must download pretrained R-101-FPN as backbone and put it in ./mmdetection/train_config folder

Run sh model.sh train for Training model

Inference

Run sh model.sh test [.pth file] for inference

.json file will be save in ./mmdetection/

you can run sh sub.sh {filename_zip} for zipping file quickly

Pre-trained model

R-101-FPN as backbone, put it in ./mmdetection/traon_config folder

Mask R-CNN-R-101-FPN is my model file, you can generate prediction result after checking Evaluation part

Results

model name	mAP
Mask R-CNN-R-101-FPN (840*840) + RandomCrop (600 * 600) + anchor scale =8	0.24518
Mask R-CNN-R-101-FPN (840*840) + RandomCrop (600 * 600) + anchor scale =6	0.244608
Mask R-CNN-R-101-FPN (640*640) + anchor scale =8	0.221494