English | 简体中文
-
Feature models:
- Object detection: The lightweight object detection model PP-PicoDet, performace and inference speed reaches SOTA on mobile side
- Keypoint detection: The lightweight keypoint detection model PP-TinyPose for mobile side
-
Model richness:
- Object detection:
- Publish Swin-Transformer object detection model
- Publish TOOD(Task-aligned One-stage Object Detection) model
- Publish GFL(Generalized Focal Loss) object detection model
- Publish Sniper optimization method for tiny object detection, supporting Faster RCNN and PP-YOLO series models
- Publish PP-YOLO optimized model PP-YOLO-EB for EdgeBoard
- Multi-object tracking:
- Publish Real-time tracking system PP-Tracking
- Publish high-precision, small-scale and lightweight model based on FairMot
- Publish real-time tracking model zoo for pedestrian, head and vehicle tracking, including scenarios such as aerial surveillance, autonomous driving, dense crowds, and tiny object tracking
- DeepSort support PP-YOLO, PP-PicoDet as object detector
- Keypoint detection:
- Publish Lite HRNet model
- Object detection:
-
Inference deployment:
- Support NPU deployment for YOLOv3 series
- Support C++ deployment for FairMot
- Support C++ and PaddleLite deployment for keypoint detection series model
-
Documents:
- Add series English documents
-
Model richness:
- Publish the Transformer test model: DETR, Deformable DETR, Sparse RCNN
- Key point test new Dark model, release Dark HRNet model
- Publish the MPII dataset HRNet keypoint detection model
- Release head and vehicle tracking vertical model
-
Model optimization:
- AlignConv optimization model was released by S2ANet, and DOTA dataset mAP was optimized to 74.0
-
Inference deployment
- Mainstream models support batch size>1 predictive deployment, including YOLOv3, PP-YOLO, Faster RCNN, SSD, TTFNet, FCOS
- New addition of target tracking models (JDE, Fair Mot, Deep Sort) Python side prediction deployment support, and support for TensorRT prediction
- FairMot joint key point detection model deployment Python side predictive deployment support
- Added support for key point detection model combined with PP-YOLO prediction deployment
-
Documents:
- New TensorRT version notes to Windows Predictive Deployment documentation
- FAQ documents are updated
-
Bug fixes:
- Fixed PP-YOLO series model training convergence problem
- Fixed the problem of no label data training when batch_size > 1
-
Model richness enhancement:
- Key point model: HRNet, HigherHRNet
- Publish the multi-target tracking model: DeepSort, FairMot, JDE
-
Basic framework Capabilities:
- Supports training without labels
-
Forecast deployment:
- Paddle Inference YOLOv3 series model support batch_size>1 prediction
- Rotating frame detection S2ANet model prediction deployment is open
- Incremental quantization model benchmark
- Add dynamic graph model and static graph model: Paddle-Lite demo
-
Detection model compression:
- Release PP-YOLO series model compression model
-
Documents:
- Update quick start, forecast deployment and other tutorial documentation
- Added ONNX model export tutorial
- Added the mobile deployment document
Description: Since version 2.0, dynamic graphs are used as the default version of Paddle Detection, the original dygraph directory is switched to the root directory, and the original static graph implementation is moved to the static directory.
-
Enhancement of dynamic graph model richness:
- PP-YOLOv2 and PP-YOLO tiny models were published. The accuracy of PP-YOLOv2 COCO Test dataset reached 49.5%, and the prediction speed of V100 reached 68.9 FPS
- Release the rotary frame detection model S2ANet
- Release the two-phase utility model PSS-Det
- Publish the face detection model Blazeface
-
New basic module:
- Added SENet, GhostNet, and Res2Net backbone networks
- Added VisualDL training visualization support
- Added single precision calculation and PR curve drawing function
- The YOLO models support THE NHWC data format
-
Forecast deployment:
- Publish forecast benchmark data for major models
- Adaptive to TensorRT6, support TensorRT dynamic size input, support TensorRT int8 quantitative prediction
- 7 types of models including PP-YOLO, YOLOv3, SSD, TTFNet, FCOS, Faster RCNN are deployed in Python/CPP/TRT prediction on Linux, Windows and NV Jetson platforms
-
Detection model compression:
- Distillation: Added dynamic map distillation support and released YOLOv3-MobileNetV1 distillation model
- Joint strategy: new dynamic graph prunning + distillation joint strategy compression scheme, and release YOLOv3-MobileNetV1 prunning + distillation compression model
- Problem fix: Fixed dynamic graph quantization model export problem
-
Documents:
- New English document of dynamic graph: including homepage document, getting started, quick start, model algorithm, new dataset, etc
- Added both English and Chinese installation documents of dynamic diagrams
- Added configuration file templates and description documents of dynamic graph RCNN series and YOLO series
-
Enhancement of dynamic graph model richness:
- Optimize networking and training mode of RCNN models, and improve accuracy of RCNN series models (depending on Paddle Develop or version 2.0.1)
- Added support for SSDLite, FCOS, TTFNet, SOLOv2 series models
- Added pedestrian and vehicle vertical object detection models
-
New dynamic graph basic module:
- Added MobileNetV3 and HRNet backbone networks
- Improved roi-align calculation logic for RCNN series models (depending on Paddle Develop or version 2.0.1)
- Added support for Synchronized Batch Norm
- Added support for Modulated Deformable Convolution
-
Forecast deployment:
- Publish dynamic diagrams in python, C++, and Serving deployment solution and documentation. Support Faster RCNN, Mask RCNN, YOLOv3, PPYOLO, SSD, TTFNet, FCOS, SOLOv2 and other models to predict deployment
- Dynamic graph prediction deployment supports TensorRT mode FP32, FP16 inference acceleration
-
Detection model compression:
- Prunning: Added dynamic graph prunning support, and released YOLOv3-MobileNetV1 prunning model
- Quantization: Added quantization support of dynamic graph, and released quantization models of YOLOv3-MobileNetV1 and YOLOv3-MobileNetV3
-
Documents:
- New Dynamic Diagram tutorial documentation: includes installation instructions, quick start, data preparation, and training/evaluation/prediction process documentation
- New advanced tutorial documentation for dynamic diagrams: includes documentation for model compression and inference deployment
- Added dynamic graph model library documentation
-
Dynamic graph support:
- Support for Faster-RCNN, Mask-RCNN, FPN, Cascade Faster/Mask RCNN, YOLOv3 and SSD models, trial version.
-
Model upgrade:
- Updated PP-YOLO Mobile-Netv3 large and small models with improved accuracy, and added prunning and distillation models.
-
New features:
- Support VisualDL visual data preprocessing pictures.
-
Bug fix:
- Fix Blaze Face keypoint prediction bug.
-
Model richness enhancement:
- SOLOv2 series models were released, in which the SOLOv2-Light-R50-VD-DCN-FPN model achieved 38.6 FPS on a single gpu V100, accelerating by 24%, and the accuracy of COCO verification set reached 38.8%, improving by 2.4 absolute percentage points.
- Added Android mobile terminal detection demo, including SSD, YOLO series model, can directly scan code installation experience.
-
Mobile terminal model optimization:
- Added to PACT's new quantization strategy, YOLOv3 Mobilenetv3 is 0.7% better than normal quantization on COCO datasets.
-
Ease of use and functional components:
- Enhance the function of generate_proposal_labels operator to avoid nan risk of the model.
- Fixed several problems with deploy python and C++ prediction.
- Unified COCO and VOC datasets under the evaluation process, support the output of a single class of AP and P-R curves.
- PP-YOLO supports rectangular input images.
-
Documents:
- Added object detection whole process tutorial, added Jetson platform deployment tutorial.
-
Model richness enhancement:
- The PPYOLO model was released. The accuracy of COCO dataset reached 45.2%, and the prediction speed of single gpu V100 reached 72.9 FPS, which was better than that of YOL Ov4 model.
- New TTFNet model, base version aligned with competing products, COCO dataset accuracy up to 32.9%.
- New HTC model, base version aligned with competing products, COCO dataset accuracy up to 42.2%.
- BlazeFace key point detection model was added, with an accuracy of 85.2% in Wider-Face's Easy-Set.
- ACFPN model was added, and the accuracy of COCO dataset reached 39.6%.
- General object detection model (including 676 classes) on the publisher side. On the COCO dataset with the same strategy, when V100 is 19.5FPS, the COCO mAP can reach 49.4%.
-
Mobile terminal model optimization:
- Added SSD Lite series optimization models, including Ghost Net Backbone, FPN components, etc., with accuracy improved by 0.5% and 1.5%.
-
Ease of use and functional components:
- Add GridMask, Random Erasing data enhancement method.
- Added support for Matrix NMS.
- EMA(Exponential Moving Average) training support.
- The new multi-machine training method, the average acceleration ratio of two machines to single machine is 80%, multi-machine training support needs to be further verified.
-
Model richness enhancement:
- Efficientdet-D0 model added, speed and accuracy is better than competing products.
- Added YOLOv4 prediction model, precision aligned with competing products; Added YOLOv4 fine tuning training on Pascal VOC datasets with accuracy of 85.5%.
- YOLOv3 added MobileNetV3 backbone network, COCO dataset accuracy reached 31.6%.
- Add Anchor-free model FCOS, the accuracy is better than competing products.
- Anchor-free model Cornernet Squeeze was added, the accuracy was better than competing products, and the accuracy of COCO dataset of optimized model was 38.2% and +3.7%, 5% faster than YOL Ov3 Darknet53.
- The CascadeRCNN-ResNet50vd model, which is a practical object detection model on the server side, is added, and its speed and accuracy are better than that of the competitive EfficientDet.
-
Mobile terminal launched three models:
- SSSDLite model: SSDLite-Mobilenetv3 small/large model, with better accuracy than competitors.
- YOLOv3 Mobile solution: The YOLOv3-MobileNetv3 model accelerates 3.5 times after compression, which is faster and more accurate than the SSD Lite model of competing products.
- RCNN Mobile terminal scheme: CascadeRCNN-MobileNetv3, after series optimization, launched models with input images of 320x320 and 640x640 respectively, with high cost performance for speed and accuracy.
-
Anticipate deployment refactoring:
- New Python prediction deployment process, support for RCNN, YOLO, SSD, Retina Net, face models, support for video prediction.
- Refactoring C++ predictive deployment to improve ease of use.
-
Ease of use and functional components:
- Added Auto Augment data enhancement.
- Upgrade the detection library document structure.
- Support shape matching automatically by transfer learning.
- Optimize memory footprint during mask branch evaluation.
- The new model:
- Added CBResNet model.
- Added LibraRCNN model.
- The accuracy of YOLOv3 model was further improved, and the accuracy based on COCO data reached 43.2%, 1.4% higher than the previous version.
- New Basic module:
- Trunk network: CBResNet is added.
- Loss module: Loss of YOLOv3 supports fine-grained OP combinations.
- Regular module: Added the Drop Block module.
- Function optimization and improvement:
- Accelerate YOLOv3 data preprocessing and increase the overall training speed by 40%.
- Optimize data preprocessing logic to improve ease of use.
- dd face detection prediction benchmark data.
- Added C++ prediction engine Python API prediction example.
- Detection model compression:
- prunning: Release MobileNet-YOLOv3 prunning scheme and model, based on VOC data FLOPs 69.6%, mAP + 1.4%, based on COCO DATA FLOPS 28.8%, mAP + 0.9%; Release ResNet50vd-DCN-YOLOv3 clipped solution and model based on COCO datasets 18.4%, mAP + 0.8%.
- Distillation: Release MobileNet-YOLOv3 distillation scheme and model, based on VOC data mAP + 2.8%, COCO data mAP + 2.1%.
- Quantification: Release quantification models of YOLOv3 Mobile Net and Blaze Face.
- Prunning + distillation: release MobileNet-YOLOv3 prunning + distillation solution and model, 69.6% based on COCO DATA FLOPS, 64.5% based on TensorRT prediction acceleration, 0.3% mAP; Release ResNet50vd-DCN-YOLOv3 tailoring + distillation solution and model, 43.7% based on COCO Data FLOPS, 24.0% based on TensorRT prediction acceleration, mAP + 0.6%.
- Search: Open source Blaze Face Nas complete search solution.
- Predict deployment:
- Integrated TensorRT, support FP16, FP32, INT8 quantitative inference acceleration.
- Document:
- Add detailed data preprocessing module to introduce documents and implement custom data Reader documents.
- Added documentation on how to add algorithm models.
- Document deployment to the web site: https://paddledetection.readthedocs.io
- Add Res2Net model.
- Add HRNet model.
- Add GIOU loss and DIOU loss。
- Add CascadeClsAware RCNN model.
- Add CBNet, ResNet200 and Non-local model.
- Add SoftNMS.
- Add Open Image V5 dataset and Objects365 dataset model
- Added enhanced YOLOv3 model with accuracy up to 41.4%.
- Added Face detection models BlazeFace and Faceboxes.
- Rich COCO based models, accuracy up to 51.9%.
- Added CA-Cascade-RCNN, one of the best single models to win on Objects365 2019 Challenge.
- Add pedestrian detection and vehicle detection pre-training models.
- Support FP16 training.
- Added cross-platform C++ inference deployment scheme.
- Add model compression examples.
- Add GroupNorm model.
- Add CascadeRCNN+Mask model.
- Add Modulated Deformable Convolution series model
- Add detection library Chinese document
- Fixed an issue where R-CNN series model training was evaluated simultaneously
- Add ResNext101-vd + Mask R-CNN + FPN models
- Added YOLOv3 model based on VOC dataset
- First release of PaddleDetection Detection library and Detection model library
- models：Faster R-CNN, Mask R-CNN, Faster R-CNN+FPN, Mask R-CNN+FPN, Cascade-Faster-RCNN+FPN, RetinaNet, YOLOv3, 和SSD.