Skip to content

Commit 06712bd

Browse files
CuberrChenCuberrChen
committed
update one card result and Readme
1 parent 940f94a commit 06712bd

File tree

4 files changed

+6976
-38
lines changed

4 files changed

+6976
-38
lines changed

‎README.md

Lines changed: 50 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -24,7 +24,7 @@ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (C
2424
## 2 复现精度
2525
>在KITTI val数据集（50/50 split as paper）的测试效果如下表。
2626
27-
1、当网络结构和损失函数以及大部分数据处理、训练配置和论文一致时，cls loss和loc loss的权重分配（论文里写的1：1,这里经过实验1：2结果更好）和batch size以及学习率不同。
27+
1、当网络结构和损失函数以及大部分数据处理、训练配置和论文一致时，batch size以及学习率不同。
2828
所能达到的结果如下表所示：
2929

3030
|NetWork |epochs|opt|lr|batch_size|dataset|config|
@@ -33,25 +33,25 @@ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (C
3333

3434
```
3535
Car [email protected], 0.70, 0.70:
36-
bbox AP:89.95, 86.19, 78.88
37-
bev AP:89.54, 85.82, 78.64
38-
3d AP:75.03, 64.79, 62.19
39-
aos AP:45.54, 43.73, 39.87
36+
bbox AP:90.26, 86.24, 79.26
37+
bev AP:89.92, 86.04, 79.14
38+
3d AP:77.00, 66.40, 63.24
39+
aos AP:38.34, 37.30, 33.19
4040
Car [email protected], 0.50, 0.50:
41-
bbox AP:89.95, 86.19, 78.88
42-
bev AP:90.66, 89.32, 88.47
43-
3d AP:90.60, 88.84, 87.71
44-
aos AP:45.54, 43.73, 39.87
41+
bbox AP:90.26, 86.24, 79.26
42+
bev AP:90.80, 89.84, 88.88
43+
3d AP:90.75, 89.32, 87.84
44+
aos AP:38.34, 37.30, 33.19
4545
4646
Car coco [email protected]:0.05:0.95:
47-
bbox AP:66.21, 62.33, 58.75
48-
bev AP:66.81, 63.10, 59.98
49-
3d AP:53.04, 48.61, 45.87
50-
aos AP:34.47, 32.44, 30.35
47+
bbox AP:67.72, 63.70, 61.10
48+
bev AP:67.13, 63.44, 61.15
49+
3d AP:53.45, 48.92, 46.34
50+
aos AP:28.82, 27.54, 25.55
5151
```
52-
预训练权重和日志：[百度网盘]() | [AiStudio存储]()
52+
预训练权重和日志：[百度网盘](https://pan.baidu.com/s/1MQ9do53CJHEjtXoD1eSTCg?pwd=fmxk) | [AiStudio存储](https://aistudio.baidu.com/aistudio/datasetdetail/124683)
5353

54-
2、当将分类损失改为FocalLoss以及加入针对aos的direction分类损失时，结果有提升。
54+
2、当将分类损失改为FocalLoss以及加入针对aos的direction分类损失时(后续实验表明direction损失只对aos起作用，可不用)
5555

5656
|NetWork |epochs|opt|lr|batch_size|dataset|config|
5757
| :---: | :---: | :---: | :---: | :---: | :---: |:---: |
@@ -86,7 +86,36 @@ aos AP:66.89, 62.19, 59.23
8686
### 1. 克隆项目
8787

8888
```bash
89-
git clone https://github.com/CuberrChen/VoxelNet.git
89+
git clone [email protected]:CuberrChen/VoxelNet.git
90+
```
91+
92+
项目结构：
93+
94+
```
95+
VoxelNet/
96+
├── images
97+
├── log
98+
├── paddleplus
99+
│ ├── nn
100+
│ ├── ops
101+
│ ├── train
102+
│ ├── __init__.py
103+
│ ├── metrics.py
104+
│ └── tools.py
105+
├── README_EN.md
106+
├── README.md
107+
├── requirements.txt
108+
└── voxelnet
109+
├── builder
110+
├── configs
111+
├── core
112+
├── data
113+
├── kittiviewer
114+
├── output
115+
├── pypaddle
116+
├── utils
117+
├── __init__.py
118+
└── create_data.py
90119
```
91120

92121
### 2. 安装依赖
@@ -130,7 +159,7 @@ export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/VoxelNet
130159

131160
* Dataset preparation
132161

133-
首先下载 [KITTI 3D目标检测的数据集](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d) 并创建一些文件夹:
162+
首先下载 [官方KITTI 3D目标检测的数据集](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d) 或者AiStudio上的数据集：[kitti_detection](https://aistudio.baidu.com/aistudio/datasetdetail/50186) 并创建一些文件夹:
134163

135164
```plain
136165
└── KITTI_DATASET_ROOT # KITTI数据集的路径
@@ -205,17 +234,17 @@ eval_input_reader: {
205234

206235
### Train
207236

237+
单卡：
208238
```bash
209239
python ./pypaddle/train.py train --config_path=./configs/config.py --model_dir=./output
210240
```
211-
241+
多卡：
212242
```
213-
python -m paddle.distributed.launch ./pypaddle/train_mgpu.py --config_path=./configs/config.py --model_dir=./output
243+
python -m paddle.distributed.launch ./pypaddle/train_mgpu.py --config_path=./configs/configFix.py --model_dir=./output
214244
```
215245
注意：
216246

217-
* batch size 2 时，训练显存大约11G。可通过修改post_center_limit_range Z的范围以及max_number_of_voxels大小节省显存。
218-
247+
* batch size 2 时，单卡训练显存大约11G。可通过修改post_center_limit_range Z的范围以及max_number_of_voxels大小节省显存。
219248

220249
### Evaluate
221250
```bash

‎README_EN.md

Lines changed: 45 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -33,25 +33,25 @@ The achieved results are shown in the following table：
3333

3434
```
3535
Car [email protected], 0.70, 0.70:
36-
bbox AP:89.95, 86.19, 78.88
37-
bev AP:89.54, 85.82, 78.64
38-
3d AP:75.03, 64.79, 62.19
39-
aos AP:45.54, 43.73, 39.87
36+
bbox AP:90.26, 86.24, 79.26
37+
bev AP:89.92, 86.04, 79.14
38+
3d AP:77.00, 66.40, 63.24
39+
aos AP:38.34, 37.30, 33.19
4040
Car [email protected], 0.50, 0.50:
41-
bbox AP:89.95, 86.19, 78.88
42-
bev AP:90.66, 89.32, 88.47
43-
3d AP:90.60, 88.84, 87.71
44-
aos AP:45.54, 43.73, 39.87
41+
bbox AP:90.26, 86.24, 79.26
42+
bev AP:90.80, 89.84, 88.88
43+
3d AP:90.75, 89.32, 87.84
44+
aos AP:38.34, 37.30, 33.19
4545
4646
Car coco [email protected]:0.05:0.95:
47-
bbox AP:66.21, 62.33, 58.75
48-
bev AP:66.81, 63.10, 59.98
49-
3d AP:53.04, 48.61, 45.87
50-
aos AP:34.47, 32.44, 30.35
47+
bbox AP:67.72, 63.70, 61.10
48+
bev AP:67.13, 63.44, 61.15
49+
3d AP:53.45, 48.92, 46.34
50+
aos AP:28.82, 27.54, 25.55
5151
```
52-
Pre-trained weights and training log：[]()
52+
Pre-trained weights and training log：[Baidu Cloud](https://pan.baidu.com/s/1MQ9do53CJHEjtXoD1eSTCg?pwd=fmxk) | [AiStudio](https://aistudio.baidu.com/aistudio/datasetdetail/124683)
5353

54-
2、The results are significantly improved when the CrossEntropy loss is changed to FocalLoss and when the direction classification loss for aos is added
54+
2、The results are improved when the CrossEntropy loss is changed to FocalLoss and when the direction classification loss for aos is added
5555

5656
|NetWork |epochs|opt|lr|batch_size|dataset|config|
5757
| :---: | :---: | :---: | :---: | :---: | :---: |:---: |
@@ -64,7 +64,7 @@ bev AP:89.79, 85.26, 78.93
6464
aos AP:89.81, 84.55, 77.71
6565
Car [email protected], 0.50, 0.50:
6666
bbox AP:90.19, 85.78, 79.38
67-
bev AP:96.51, 89.53, 88.59
67+
bev AP:96.51, 89.53, 88.59
6868
3d AP:90.65, 89.08, 87.52
6969
aos AP:89.81, 84.55, 77.71
7070
@@ -76,7 +76,7 @@ aos AP:66.89, 62.19, 59.23
7676
```
7777
Pre-trained weights and training log：[Baidu Cloud](https://pan.baidu.com/s/1LuB5N_CbzWT5HyFDm-a66g?pwd=3633) | [AiStudio](https://aistudio.baidu.com/aistudio/datasetdetail/124650)
7878

79-
**In addition, the details not mentioned in the paper, this project are referred to the implementation of the Second project**
79+
**In addition, this project are referred to the implementation of the Second project for the details not mentioned in the paper, **
8080

8181
## 3 Start
8282

@@ -85,6 +85,34 @@ Pre-trained weights and training log：[Baidu Cloud](https://pan.baidu.com/s/1Lu
8585
```bash
8686
git clone https://github.com/CuberrChen/VoxelNet.git
8787
```
88+
project structure:
89+
```
90+
VoxelNet/
91+
├── images
92+
├── log
93+
├── paddleplus
94+
│ ├── nn
95+
│ ├── ops
96+
│ ├── train
97+
│ ├── __init__.py
98+
│ ├── metrics.py
99+
│ └── tools.py
100+
├── README_EN.md
101+
├── README.md
102+
├── requirements.txt
103+
└── voxelnet
104+
├── builder
105+
├── configs
106+
├── core
107+
├── data
108+
├── kittiviewer
109+
├── output
110+
├── pypaddle
111+
├── utils
112+
├── __init__.py
113+
└── create_data.py
114+
```
115+
88116

89117
### 2. Dependencies
90118
The most suitable environment configuration：
@@ -126,7 +154,7 @@ export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/VoxelNet
126154

127155
* Dataset preparation
128156

129-
Fristly, Download [KITTI 3D Object Det](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d) and create some folders:
157+
Fristly, Download Official [KITTI 3D Object Det](http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d) or AiStudio [kitti_detection](https://aistudio.baidu.com/aistudio/datasetdetail/50186) and create some folders:
130158

131159
```plain
132160
└── KITTI_DATASET_ROOT

0 commit comments

Comments
 (0)