Data Preparation Example

[blueclowd]

Just to share my data preparation process. Hopefully it is helpful to anyone trying this work for the first time.

Using dataset ss360 as an example

1. Download dataset from https://drive.google.com/file/d/1myN8X48Yw49fR-na2tYpzI6sxcoLlA4w/view?usp=drive_link

2. Unzip ss360_raw.zip. You will get ss360_raw consisting of

   • image/
   • train.json
   • val.json
     (Some example set like york_raw doesn't have train.json which cause errors)

3. Copy the folder ss360_raw to {project}/dataset. The {project}/dataset now consists of

   • ss360_raw/
   • json2npz.py
   • json2npt_gt.py

4. Run python json2npz.py --dataset_name ss360 (The dataset name is ss360 instead of ss360_raw)

5. Run python json2npz_gt.py --dataset_name ss360 (The dataset name is ss360 instead of ss360_raw)

6. Two more folders generated under {project}/dataset which now has

   • ss360/
   • ss360_1/
   • ss360_raw/
   • json2npz.py
   • json2npt_gt.py

7. Go back to project folder and start the training with any model name you like (its not neural network name but any name meaningful to you)
   cd ..
python train.py --dataset_name ss360 --order 1 --model_name test