Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
You can not select more than 25 topics
Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
JessonGuo
32b7edf4b1
|
1 year ago | |
|---|---|---|
| configs | 1 year ago | |
| experiments/car | 1 year ago | |
| experiments_t | 1 year ago | |
| rank_0 | 1 year ago | |
| result | 1 year ago | |
| scripts | 1 year ago | |
| src | 1 year ago | |
| train_output | 1 year ago | |
| README.md | 1 year ago | |
| __init__.py | 1 year ago | |
| eval.py | 1 year ago | |
| export.py | 1 year ago | |
| requirements.txt | 1 year ago | |
| test.py | 1 year ago | |
| train.py | 1 year ago | |
| train0.py | 1 year ago | |
| train_modelarts.py | 1 year ago | |
README.md
Contents
- Contents
PointPillars description
PointPillars is a method for object detection in 3D that enables end-to-end learning with only 2D convolutional layers.
PointPillars uses a novel encoder that learn features on pillars (vertical columns) of the point cloud to predict 3D oriented boxes for objects.
There are several advantages of this approach.
First, by learning features instead of relying on fixed encoders, PointPillars can leverage the full information represented by the point cloud.
Further, by operating on pillars instead of voxels there is no need to tune the binning of the vertical direction by hand.
Finally, pillars are highly efficient because all key operations can be formulated as 2D convolutions which are extremely efficient to compute on a GPU.
An additional benefit of learning features is that PointPillars requires no hand-tuning to use different point cloud configurations.
For example, it can easily incorporate multiple lidar scans, or even radar point clouds.
Paper: PointPillars: Fast Encoders for Object Detection from Point Clouds.
Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, Oscar Beijbom, 2018.
Model architecture
The main components of the network are a Pillar Feature Network, Backbone, and SSD detection head.
The raw point cloud is converted to a stacked pillar tensor and pillar index tensor.
The encoder uses the stacked pillars to learn a set of features that can be scattered back to a 2D pseudo-image for a convolutional neural network.
The features from the backbone are used by the detection head to predict 3D bounding boxes for objects.
Dataset
Dataset used: KITTI
Data was collected with using a standard station wagon with two high-resolution color and grayscale video cameras.
Accurate ground truth is provided by a Velodyne laser scanner and a GPS localization system.
Dataset was captured by driving around the mid-size city of Karlsruhe, in rural areas and on highways.
Up to 15 cars and 30 pedestrians are visible per image. The 3D object detection benchmark consists of 7481 images.
Environment requirements
-
Hardware（GPU）
- Prepare hardware environment with GPU processor.
-
Framework
-
For more information, please check the resources below：
-
Download KITTI, data from ImageSets, put files from
ImageSetsintopointpillars/src/data/ImageSets/For the dataset preparation instructions see the Dataset preparation section.
Quick start
After preparing the dataset you can start training and evaluation as follows：
Running on GPU
Train
# standalone train
bash ./scripts/run_standalone_train_gpu.sh [CFG_PATH] [SAVE_PATH] [DEVICE_ID]
# distribute train
bash ./scripts/run_distribute_train_gpu.sh [CFG_PATH] [SAVE_PATH] [DEVICE_NUM]
Example:
# standalone train
bash ./scripts/run_standalone_train_gpu.sh ./configs/car_xyres16.yaml ./experiments/car/ 0
# distribute train (8p)
bash ./scripts/run_distribute_train_gpu.sh ./configs/car_xyres16.yaml ./experiments/car/ 8
Evaluate
# evaluate
bash ./scripts/run_eval_gpu.sh [CFG_PATH] [CKPT_PATH] [DEVICE_ID]
Example:
# evaluate
bash ./scripts/run_eval_gpu.sh ./configs/car_xyres16.yaml ./experiments/car/poinpitllars-160_37120.ckpt 0
Script Description
Script and Sample Code
.
└── pointpillars
├── configs
│ ├── car_xyres16.yaml # config for car detection
│ └── ped_cycle_xyres16.yaml # config for pedestrian and cyclist detection
├── scripts
│ ├── run_distribute_train_gpu.sh # launch distributed training(8p) on GPU
│ ├── run_eval_gpu.sh # launch evaluating on GPU
│ ├── run_export_gpu.sh # launch export mindspore model to mindir
│ └── run_standalone_train_gpu.sh # launch standalone traininng(1p) on GPU
├── src
│ ├── builder
│ │ ├── __init__.py # init file
│ │ ├── anchor_generator_builder.py # builder for anchor generator
│ │ ├── box_coder_builder.py # builder for box coder
│ │ ├── dataset_builder.py # builder for dataset
│ │ ├── dbsampler_builder.py # builder for db sampler
│ │ ├── model_builder.py # builder for model
│ │ ├── preprocess_builder.py # builder for preprocess
│ │ ├── similarity_calculator_builder.py # builder for similarity calculator
│ │ ├── target_assigner_builder.py # builder for target assigner
│ │ └── voxel_builder.py # builder for voxel generator
│ ├── core
│ │ ├── point_cloud
│ │ │ ├── __init__.py # init file
│ │ │ ├── bev_ops.py # ops for bev
│ │ │ └── point_cloud_ops.py # ops for point clouds
│ │ ├── __init__.py # init file
│ │ ├── anchor_generator.py # anchor generator
│ │ ├── box_coders.py # box coders
│ │ ├── box_np_ops.py # box ops with numpy
│ │ ├── box_ops.py # box ops with mindspore
│ │ ├── einsum.py # einstein sum
│ │ ├── eval_utils.py # utils for evaluate
│ │ ├── geometry.py # geometry
│ │ ├── losses.py # losses
│ │ ├── nms.py # nms
│ │ ├── preprocess.py # preprocess operations
│ │ ├── region_similarity.py # region similarity calculator
│ │ ├── sample_ops.py # ops for sample data
│ │ ├── target_assigner.py # target assigner
│ │ └── voxel_generator.py # voxel generator
│ ├── data
│ │ ├── ImageSets # splits for train and val
│ │ ├── __init__.py # init file
│ │ ├── dataset.py # kitti dataset
│ │ ├── kitti_common.py # auxiliary file for kitti
│ │ └── preprocess.py # preprocess dataset
│ ├── create_data.py # create dataset for train model
│ ├── __init__.py # init file
│ ├── pointpillars.py # pointpillars model
│ └── predict.py # postprocessing pointpillars`s output
├── __init__.py # init file
├── eval.py # evaluate mindspore model
├── export.py # convert mindspore model to mindir
├── README.md # readme file
├── requirements.txt # requirements
└── train.py # train mindspore model
Script Parameters
Training parameters can be configured in car_xyres16.yaml for car detection or ped_cycle_xyres16.yaml for pedestrian
and cyclist detection.
"initial_learning_rate": 0.0002, # learning rate
"max_num_epochs": 160, # number of training epochs
"weight_decay": 0.0001, # weight decay
"batch_size": 2, # batch size
"max_number_of_voxels": 12000, # mux number of voxels in one pillar
"max_number_of_points_per_voxel": 100 # max number of points per voxel
For more parameters refer to the contents of car_xyres16.yaml or ped_cycle_xyres16.yaml.
Dataset Preparation
- Add
/path/to/pointpillars/to yourPYTHONPATH
export PYTHONPATH=/path/to/pointpillars/:$PYTHONPATH
- Download KITTI dataset into one folder:
- Download left color images of object data set (12 GB)
- Download camera calibration matrices of object data set (16 MB)
- Download training labels of object data set (5 MB)
- Download Velodyne point clouds, if you want to use laser information (29 GB)
- Unzip all downloaded archives.
- Directory structure is as follows:
└── KITTI
├── training
│ ├── image_2 <-- for visualization
│ ├── calib
│ ├── label_2
│ ├── velodyne
│ └── velodyne_reduced <-- create this empty directory
└── testing
├── image_2 <-- for visualization
├── calib
├── velodyne
└── velodyne_reduced <-- create this empty directory
-
Download ImageSets, put files from
ImageSetsintopointpillars/src/data/ImageSets/ -
Create KITTI infos:
python create_data.py create_kitti_info_file --data_path=KITTI_DATASET_ROOT
- Create reduced point cloud:
python create_data.py create_reduced_point_cloud --data_path=KITTI_DATASET_ROOT
- Create groundtruth-database infos:
python create_data.py create_groundtruth_database --data_path=KITTI_DATASET_ROOT
- The config file
car_xyres16.yamlorped_cycle_xyres16.yamlneeds to be edited to point to the above datasets:
train_input_reader:
...
database_sampler:
database_info_path: "/path/to/kitti_dbinfos_train.pkl"
...
kitti_info_path: "/path/to/kitti_infos_train.pkl"
kitti_root_path: "KITTI_DATASET_ROOT"
...
eval_input_reader:
...
kitti_info_path: "/path/to/kitti_infos_val.pkl"
kitti_root_path: "KITTI_DATASET_ROOT"
Training Process
Run on GPU
Standalone training
bash ./scripts/run_standalone_train_gpu.sh ./configs/car_xyres16.yaml ./experiments/cars/ 0
Logs will be saved to ./experiments/cars/log.txt
Result:
2022-01-18 11:29:13 epoch:0, iter:1000, loss:1.1359277, fps:14.38 imgs/sec
2022-01-18 11:29:20 epoch:0, iter:1050, loss:1.0492299, fps:13.57 imgs/sec
2022-01-18 11:29:27 epoch:0, iter:1100, loss:0.884439, fps:14.24 imgs/sec
2022-01-18 11:29:34 epoch:0, iter:1150, loss:1.0804198, fps:14.34 imgs/sec
2022-01-18 11:29:41 epoch:0, iter:1200, loss:0.92863345, fps:14.19 imgs/sec
2022-01-18 11:29:49 epoch:0, iter:1250, loss:0.8504363, fps:13.51 imgs/sec
2022-01-18 11:29:56 epoch:0, iter:1300, loss:1.0816091, fps:13.89 imgs/sec
2022-01-18 11:30:03 epoch:0, iter:1350, loss:0.98323077, fps:13.51 imgs/sec
2022-01-18 11:30:10 epoch:0, iter:1400, loss:0.824274, fps:14.33 imgs/sec
2022-01-18 11:30:18 epoch:0, iter:1450, loss:0.9153076, fps:13.77 imgs/sec
Distribute training (8p)
bash ./scripts/run_distribute_train_gpu.sh ./confgs/car_xyres16.yaml ./experiments/cars/ 8
Logs will be saved to ./experiments/cars/log.txt
Result:
2022-01-18 08:30:39 epoch:1, iter:300, loss:0.9105564, fps:66.5 imgs/sec
2022-01-18 08:30:51 epoch:1, iter:350, loss:1.0566418, fps:68.85 imgs/sec
2022-01-18 08:31:03 epoch:1, iter:400, loss:0.98004365, fps:65.54 imgs/sec
2022-01-18 08:31:14 epoch:1, iter:450, loss:0.9133666, fps:69.86 imgs/sec
2022-01-18 08:31:27 epoch:2, iter:500, loss:0.8083154, fps:64.09 imgs/sec
2022-01-18 08:31:39 epoch:2, iter:550, loss:0.75948864, fps:65.78 imgs/sec
2022-01-18 08:31:51 epoch:2, iter:600, loss:1.0096964, fps:66.61 imgs/sec
2022-01-18 08:32:03 epoch:2, iter:650, loss:0.86279136, fps:66.55 imgs/sec
2022-01-18 08:32:15 epoch:3, iter:700, loss:0.89273417, fps:65.99 imgs/sec
2022-01-18 08:32:27 epoch:3, iter:750, loss:0.90943766, fps:67.59 imgs/sec
Evaluation Process
GPU
bash ./scripts/run_eval_gpu.sh [CFG_PATH] [CKPT_PATH] [DEVICE_ID]
Example:
bash ./scripts/run_eval_gpu.sh ./configs/car_xyres16.yaml ./experiments/car/pointpillars-160_37120.ckpt 0
Result:
Here is model for cars detection as an example，you can view the result in log file ./experiments/car/log_eval.txt：
Easy Mod Hard
Car AP@0.70, 0.70, 0.70:
bbox AP:90.71, 89.17, 87.81
bev AP:89.81, 86.99, 84.99
3d AP:86.90, 76.88, 72.76
aos AP:90.43, 88.47, 86.77
Here is result for pedestrian and cyclist detection：
Easy Mod Hard
Cyclist AP@0.50, 0.50, 0.50:
bbox AP:84.63, 65.24, 62.17
bev AP:83.25, 63.01, 59.65
3d AP:80.13, 60.26, 56.07
aos AP:83.79, 63.86, 60.78
Pedestrian AP@0.50, 0.50, 0.50:
bbox AP:66.80, 63.98, 60.59
bev AP:72.15, 67.50, 62.51
3d AP:66.48, 60.61, 55.87
aos AP:46.26, 44.97, 43.03
Mod. mAP is calculated as mean of metric values for corresponding rows of each classes.
For example mod. mAP for BEV detection benchmark is calculated as (86.99 + 63.01 + 67.50) / 3
So mod. mAP for BEV detection benchmark is 72.5, mod. mAP for 3D detection benchmark is 65.91
Also, the article contains information about splits for experimental studies and for test submission,
but authors provided splits only for experimental studies.
Results in the article are presented for test submission splits.
The samples are originally divided into 7481 training and 7518 testing samples.
For experimental studies we split the official training into 3712 training samples
and 3769 validation samples, while for our test submission we created a minival
set of 784 samples from the validation set and trained on the remaining 6733 samples.
Export MINDIR
If you want to infer the network on Ascend 310, you should convert the model to MINDIR.
GPU
bash ./scripts/run_export_gpu.sh [CFG_PATH] [CKPT_PATH] [FILE_NAME] [DEVICE_ID] [FILE_FORMAT](optional)
- CFG_PATH - path to the configuration file
- CKPT_PATH - path to the checkpoint file, containing weights of a trained Pointpillars model
- FILE_NAME - name of the output file for the exported model.
- DEVICE_ID - ID of the device, which will be used for the export.
- FILE_FORMAT (optional) - format of the exported model. ("AIR" or "MINDIR"). Default: "MINDIR".
Example:
bash ./scripts/run_export_gpu.sh ./configs/car_xyres16.yaml ./experiments/car/pointpillars.ckpt pointpillars 0
or
bash ./scripts/run_export_gpu.sh ./configs/car_xyres16.yaml ./experiments/car/pointpillars.ckpt pointpillars 0 MINDIR
The output should look as following:
pointpillars.mindir exported successfully!
Model Description
Training Performance on GPU
| Parameter | PointPillars (1p) | PointPillars (8p) | ||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Resource | 1x Nvidia RTX 3090 | 8x Nvidia RTX 3090 | ||||||||||||||||||||||||||||||||||||||||
| Uploaded date | - | - | ||||||||||||||||||||||||||||||||||||||||
| Mindspore version | 1.5.0 | 1.5.0 | ||||||||||||||||||||||||||||||||||||||||
| Dataset | KITTI | KITTI | ||||||||||||||||||||||||||||||||||||||||
| Training parameters | epoch=160, lr=0.0002, weight_decay=0.0001, batch_size=2 | epoch=160, lr=0.0016, weight_decay=0.0001, batch_size=2 | ||||||||||||||||||||||||||||||||||||||||
| Optimizer | Adam | Adam | ||||||||||||||||||||||||||||||||||||||||
| Loss function | WeightedSmoothL1LocalizationLoss, WeightedSoftmaxClassificationLoss, SigmoidFocalClassificationLoss | WeightedSmoothL1LocalizationLoss, WeightedSoftmaxClassificationLoss, SigmoidFocalClassificationLoss | ||||||||||||||||||||||||||||||||||||||||
| Speed | Car: 14 img/s; Pedestrian + Cyclist: 11 img/s | Car: 66 img/s; Pedestrian + Cyclist: 64 img/s | ||||||||||||||||||||||||||||||||||||||||
| mAP (BEV detection) |
|
|
||||||||||||||||||||||||||||||||||||||||
| mAP (3D detection) |
|
|
||||||||||||||||||||||||||||||||||||||||
| mAOS |
|
|
For cars required a bounding box overlap of 70%, while for pedestrians and cyclists required a bounding box overlap of 50%.
Difficulties are defined as follows:
- Easy: Min. bounding box height: 40 Px, Max. occlusion level: Fully visible, Max. truncation: 15 %
- Moderate: Min. bounding box height: 25 Px, Max. occlusion level: Partly occluded, Max. truncation: 30 %
- Hard: Min. bounding box height: 25 Px, Max. occlusion level: Difficult to see, Max. truncation: 50 %
Description of Random Situation
train.py script use mindspore.set_seed() to set global random seed, which can be modified.
ModelZoo Homepage
Please visit the official website homepage.
No Description
Text Python Shell C