for in-class competition task1 & task2
curated by wx. zhang
for more details refer to the pdf report
Results:
Environment requirements:
- Python 3.6
- packages: mxnet, gluonbook, gluoncv, pandas, ...
- preferably mxnet-cu92 to use CUDA accerleration (mxnet-cu90 is recommended if you also use tf)
- if not using GPU, change all ctx to mx.cpu()
Directory strcuture:
root
├── data
│ ├── trainval
│ ├── 0cec3d1f-544c-4146-8632-84b1f9fe89d3
│ ├── (image, bbox, cloud, proj files)
│ ├── ...
│ ├── test
│ ├── 0ff0a23e-5f50-4461-8ccf-2b71bead2e8e
│ ├── (image, cloud, proj files)
│ ├── ...
│ ├── train.rec (7000 512*384 images + bbox labels)
│ ├── train.lst (bbox labels and filenames)
│ ├── valid.rec (573 512*384 images + bbox labels)
│ ├── valid.lst (bbox labels and filenames)
│ ├── train_v3.rec (7000 768*432 images + class labels)
│ ├── train_v3.lst (class labels and filenames)
│ ├── valid_v3.rec (573 768*432 images + class labels)
│ ├── valid_v3.lst (class labels and filenames)
│ ├── test_v3.rec (2631 768*432 images)
│ ├── test_v3.lst (a list of filenames)
│ ├── xyz_train.txt (XYZ coordinates training data)
│ ├── xyz_valid.txt (XYZ coordinates training data)
│ │
│ ├── classes.csv (class 0,1,2)
│ ├── classes_v2.csv (class 0,1,2,3,4,5)
├── mxnet-dbc
├── ...
├── mxnet-mlp
├── ...
├── mxnet-ssd (originally forked from https://github.com/zhreshold/mxnet-ssd)
├── ...
├── params
├── (network parameters for ssd and frcnn)
├── output
├── (txt files for kaggle upload)
├── README.md
(Note: .rec file is the mxnet RecordIO format, refer to https://github.com/leocvml/mxnet-im2rec_tutorial on how to generate them)
for task1:
- train SSD: python mxnet-ssd/gta5_train.py
- predict with SSD: python mxnet-ssd/gta5_predict.py 42
- train Inception3 classifier: python mxnet-dbc/train.py
- predict with Inception3 classifier: python mxnet-dbc/predict.py
for task2:
- train MLP: python mxnet-mlp/mlp_train.py
- (after predicting with SSD) predict with MLP: python mxnet-mlp/predict.py