# fpointnet

**Repository Path**: nj_wang_wenbo/fpointnet

## Basic Information

- **Project Name**: fpointnet
- **Description**: frustum-pointnets
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-10-24
- **Last Updated**: 2022-10-26

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Contents

- [Contents](#contents)
- [frustum_pointnets_mindspore](#frustum_pointnets_mindspore)
- [Requirements](#Requirements)
- [Prepare_Training_Data](#Prepare_Training_Data)
- [Project_structure](#Project_structure)
- [Train](#train)
- [Test](#Test)

<!-- /TOC -->
# frustum_pointnets_mindspore

A mindspore version of [frustum-pointnets](https://github.com/charlesq34/frustum-pointnets)

main function of frustum_pointnets:

```angular2
train.py,
eval.py,
train/provider.py,
src/frustum_pointnets_v1.py
src/model_util.py
kitti/*
```

# Requirements

Ubuntu-18.04
CUDA-11
mindspore 1.8
python 3.9

# Prepare_Training_Data

KITTI Dataset：http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark

the sub-sets of data Of KITTI：
image_2:https://s3.eu-central-1.amazonaws.com/avg-kitti/data_object_image_2.zip

label_2:https://s3.eu-central-1.amazonaws.com/avg-kitti/data_object_label_2.zip

velodyne:https://s3.eu-central-1.amazonaws.com/avg-kitti/data_object_velodyne.zip

calib:https://s3.eu-central-1.amazonaws.com/avg-kitti/data_object_calib.zip

Download KITTI 3D object detection data and organize the folders as follows:

```angular2
.
├── dataset
│   ├── KITTI
│   │   ├── rgb_detections
│   │   │   ├──training
│   │   │   |  ├──calib
|   |   |   |  ├──velodyne
|   |   |   |  ├──label_2
|   |   |   |  ├──image_2
|   |   |   |  ├──(optional: planes)
│   │   │   ├──testing
│   │   │   |  ├──calib
|   |   |   |  ├──velodyne
|   |   |   |  ├──label_2
|   |   |   |  ├──image_2
│
```

Generate training samples from raw files:

```angular2
python kitti/prepare_data.py --gen_train --gen_val --gen_val_rgb_detection --car_only
```

or

Download the [packaged](https://pan.baidu.com/s/1uNq91d1klrweMa3eSOSdNw?pwd=fm5d) file and put the pickle files under /dataset:

# Project_structure

```angular2
.
│  eval.py
│  rank_table_2pcs.json
│  README.md
│  train.py
│
├─kitti
│  │  kitti_object.py
│  │  kitti_util.py
│  │  prepare_data.py
│  │
│  ├─image_sets
│  │      test.txt
│  │      train.txt
│  │      trainval.txt
│  │      val.txt
│  │
│  └─rgb_detections
│          rgb_detection_train.txt
│          rgb_detection_val.txt
│
├─scripts
│      run_distribution_ascend.sh
│
├─src
│      datautil.py
│      frustum_pointnets_v1.py
│      model_util.py
│
└─train
        box_util.py
        convert.py
        provider.py
```

# Train

```bash
step1: bash scripts/run_distribution_ascend.sh [RANK_TABLE_FILE]
step2: python eval.py --target 'keep_one'
```

## Train log step1

```angular2
start train ...
init environment ...
device_num:8, rank_id:0
2022-10-19T09:26:37.721182:loading train dataset ...
2022-10-19T09:26:41.325025:construct net ...
epoch: 1 step: 100, loss is 46.78062438964844
epoch: 1 step: 200, loss is 18.396862030029297
Train epoch time: 168754.465 ms, per step time: 730.539 ms
epoch: 2 step: 69, loss is 26.49061393737793
epoch: 2 step: 169, loss is 19.57318878173828
Train epoch time: 113103.595 ms, per step time: 489.626 ms
...
.
.
.
...
train down!
```

## Train log step2

```angular2
Acc:8.84  Delete:log/Fpointnet_20221019_0914/fpoint_v1-2_231.ckpt
Acc:38.46  Delete:log/Fpointnet_20221019_0914/fpoint_v1-4_231.ckpt
Acc:15.52  Delete:log/Fpointnet_20221019_0914/fpoint_v1-3_231.ckpt
Acc:16.65  Delete:log/Fpointnet_20221019_0914/fpoint_v1-1_231.ckpt
Acc:67.71  Delete:log/Fpointnet_20221019_0926/fpoint_v1-163_231.ckpt
Acc:27.84  Delete:log/Fpointnet_20221019_0926/fpoint_v1-2_231.ckpt
...
.
.
.
```

# Test

```bash
python eval.py --model_path [CKPT_PATH]
```

## Test log

```angular2
Number of point clouds: 12538
segmentation accuracy 0.8969611287934679
box IoU(ground) 0.7755848700054374
box IoU(3D) 0.7211546932116405
box estimation accuracy (IoU=0.7) 0.7359036528951986
Average pos ratio: 0.436471
Average pos prediction ratio: 0.462065
Average npoints: 1024.000000
Mean points: x0.025841 y0.987336 z24.954839
Max points: x17.198647 y9.372097 z79.747406
Min points: x-20.042944 y-3.882158 z0.000000
{'Average pos ratio:': 0.43647078506340725, 'Average pos prediction ratio:': 0.46206524353465467, 'Average npoints:': 1024.0, 'Mean points:': array([ 0.02584073,  0.98733621, 24.95483863]), 'Max points:': array([17.19864655,  9.37209702, 79.74740601]), 'Min points:': array([-20.04294395,  -3.8821578 ,   0.        ])}

```