-[Deployment for Top-Down models](#deployment-for-top-down-models)
-[Deployment for Top-Down models](#deployment-for-top-down-models)
-[Deployment for Bottom-Up models](#deployment-for-bottom-up-models)
-[Deployment for Bottom-Up models](#deployment-for-bottom-up-models)
-[joint inference with Multi-Object Tracking model FairMOT](#joint-inference-with-multi-object-tracking-model-fairmot)
-[Joint Inference with Multi-Object Tracking Model FairMOT](#joint-inference-with-multi-object-tracking-model-fairmot)
-[Train with custom data](#Train-with-custom-data)
-[BenchMark](#benchmark)
## Introduction
## Introduction
The keypoint detection part in PaddleDetection follows the state-of-the-art algorithm closely, including Top-Down and Bottom-Up methods, which can satisfy the different needs of users.
The keypoint detection part in PaddleDetection follows the state-of-the-art algorithm closely, including Top-Down and Bottom-Up methods, which can satisfy the different needs of users.
Top-Down detects the object first and then detect the specific keypoint. The accuracy of Top-Down models will be higher, but the time required will increase by the number of objects.
Top-Down detects the object first and then detect the specific keypoint. The accuracy of Top-Down models will be higher, but the time required will increase by the number of objects.
Differently, Bottom-Up detects the point first and then group or connect those points to form several instances of human pose. The speed of Bottom-Up is fixed and will not increase by the number of objects, but the accuracy will be lower.
Differently, Bottom-Up detects the point first and then group or connect those points to form several instances of human pose. The speed of Bottom-Up is fixed and will not increase by the number of objects, but the accuracy will be lower.
...
@@ -36,7 +37,7 @@ At the same time, PaddleDetection provides [PP-TinyPose](./tiny_pose/README.md)
...
@@ -36,7 +37,7 @@ At the same time, PaddleDetection provides [PP-TinyPose](./tiny_pose/README.md)
## Model Recommendation
## Model Recommendation
### Mobile Terminal
### Mobile Terminal
|Detection Model| Keypoint Model | Input Size | Accuracy of COCO| Average Inference Time (FP16) | Model Weight | Paddle-Lite Inference Model（FP16)|
|Detection Model| Keypoint Model | Input Size | Accuracy of COCO| Average Inference Time (FP16) | Model Weight | Paddle-Lite Inference Model（FP16)|
To export MOT model, please refer to [Here](../../configs/mot/README_en.md).
To export MOT model, please refer to [Here](../../configs/mot/README_en.md).
### 4.Complete Deploy Instruction and Demo
We provide standalone deploy of PaddleInference(Server-GPU)、PaddleLite(mobile、ARM)、Third-Engine(MNN、OpenVino), which is independent of training codes。For detail, please click [Deploy-docs](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/README_en.md)。
## Train with custom data
We take an example of [tinypose_256x192](.tiny_pose/README_en.md) to show how to train with custom data.
num_joints: &num_joints 17 #the number of joints in your job
train_height: &train_height 256 #the height of model input
train_width: &train_width 192 #the width of model input
hmsize: &hmsize [48, 64] #the shape of model output，usually 1/4 of [w,h]
flip_perm: &flip_perm [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]] #the correspondence between left and right keypoint id，used for flip transform。You can add an line(by "flip: False") behind of flip_pairs in RandomFlipHalfBodyTransform of TrainReader if you don't need it
num_joints_half_body: 8 #The joint numbers of half body, used for half_body transform
prob_half_body: 0.3 #The probility of half_body transform, set to 0 if you don't need it
upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] #The joint ids of half(upper) body, used to get the upper joints in half_body transform
```
For more configs, please refer to [KeyPointConfigGuide](../../docs/tutorials/KeyPointConfigGuide_en.md)。
#### 2、Others(used for test and visualization)
- In keypoint_utils.py, please set: "sigmas = np.array([.26, .25, .25, .35, .35, .79, .79, .72, .72, .62, .62, 1.07, 1.07,.87, .87, .89, .89]) / 10.0", the value indicate the variance of a joint locations，normally 0.25-0.5 means the location is highly accuracy，for example: eyes。0.5-1.0 means the location is not sure so much，for example: shoulder。0.75 is recommand if you not sure。
- In visualizer.py, please set "EDGES" in draw_pose function，this indicate the line to show between joints for visualization。
- In pycocotools you installed, please set "sigmas"，it is the same as that in keypoint_utils.py, but used for coco evaluation。
#### 3、Note for data preparation
- The data should has the same format as Coco data, and the keypoints(Nx3) and bbox(N) should be annotated.
- please set "area">0 in annotations files otherwise it will be skiped while training.
## BenchMark
We provide benchmarks in different runtime environments for your reference when choosing models. See [Keypoint Inference Benchmark](https://github.com/PaddlePaddle/PaddleDetection/blob/develop/configs/keypoint/KeypointBenchmark.md) for details.
@@ -43,8 +43,8 @@ If you want to deploy it on the mobile devives, you also need:
...
@@ -43,8 +43,8 @@ If you want to deploy it on the mobile devives, you also need:
### Pedestrian Detection Model
### Pedestrian Detection Model
| Model | Input Size | mAP (COCO Val) | Average Inference Time (FP32)| Average Inference Time (FP16) | Config | Model Weights | Deployment Model | Paddle-Lite Model（FP32) | Paddle-Lite Model（FP16)|
| Model | Input Size | mAP (COCO Val) | Average Inference Time (FP32)| Average Inference Time (FP16) | Config | Model Weights | Deployment Model | Paddle-Lite Model（FP32) | Paddle-Lite Model（FP16)|
