This issue is used to identify which onnx models can't be inferenced or need to be converted manually for Caffe removal. The related tests will be disabled temporarily until the following models can be supported.

SSD (opset 10, 12)

• resolved

• Source: https://github.com/onnx/models/tree/main/validated/vision/object_detection_segmentation/ssd/model

• Failure:
  opset-10

  [ INFO:0@0.530] global onnx_importer.cpp:3297 parseCustomLayer DNN/ONNX: unknown node type, try using custom handler for node with 5 inputs and 1 outputs: [NonMaxSuppression]:(onnx_node!NonMaxSuppression_683)
  [ INFO:0@0.530] global onnx_importer.cpp:1036 handleNode     Input[0] = 'Concat_659'
  [ INFO:0@0.530] global onnx_importer.cpp:1036 handleNode     Input[1] = 'Slice_676'
  [ INFO:0@0.530] global onnx_importer.cpp:1036 handleNode     Input[2] = 'ConstantOfShape_678'
  [ INFO:0@0.530] global onnx_importer.cpp:1036 handleNode     Input[3] = 'ConstantOfShape_680'
  [ INFO:0@0.530] global onnx_importer.cpp:1036 handleNode     Input[4] = 'ConstantOfShape_682'
  [ INFO:0@0.530] global onnx_importer.cpp:1040 handleNode     Output[0] = 'NonMaxSuppression_683'
  OpenCV(5.0.0-pre) Error: Unspecified error (Can't create layer "onnx_node!NonMaxSuppression_683" of type "NonMaxSuppression") in getLayerInstance, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/net_impl.hpp, line 106
  [ERROR:0@0.530] global onnx_importer.cpp:1032 handleNode DNN/ONNX: ERROR during processing node with 5 inputs and 1 outputs: [NonMaxSuppression]:(onnx_node!NonMaxSuppression_683) from domain='ai.onnx'
  OpenCV(5.0.0-pre) Error: Unspecified error (> Node [NonMaxSuppression@ai.onnx]:(onnx_node!NonMaxSuppression_683) parse error: OpenCV(5.0.0-pre) /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/net_impl.hpp:106: error: (-2:Unspecified error) Can't create layer "onnx_node!NonMaxSuppression_683" of type "NonMaxSuppression" in function 'getLayerInstance'
  > ) in handleNode, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/onnx/onnx_importer.cpp, line 1051

• opset-12

  OpenCV(5.0.0-pre) Error: Unspecified error (Can't create layer "onnx_node!NonMaxSuppression_683" of type "NonMaxSuppression") in getLayerInstance, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/net_impl.hpp, line 106
  [ERROR:0@0.233] global onnx_importer.cpp:1032 handleNode DNN/ONNX: ERROR during processing node with 5 inputs and 1 outputs: [NonMaxSuppression]:(onnx_node!NonMaxSuppression_683) from domain='ai.onnx'
  [ INFO:0@0.233] global onnx_importer.cpp:1036 handleNode     Input[0] = 'Concat_659'
  OpenCV(5.0.0-pre) Error: Unspecified error (> Node [NonMaxSuppression@ai.onnx]:(onnx_node!NonMaxSuppression_683) parse error: OpenCV(5.0.0-pre) /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/net_impl.hpp:106: error: (-2:Unspecified error) Can't create layer "onnx_node!NonMaxSuppression_683" of type "NonMaxSuppression" in function 'getLayerInstance'
  [ INFO:0@0.233] global onnx_importer.cpp:1036 handleNode     Input[1] = 'Slice_676'
  > ) in handleNode, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/onnx/onnx_importer.cpp, line 1051
  [ INFO:0@0.233] global onnx_importer.cpp:1036 handleNode     Input[2] = 'ConstantOfShape_678'
  [ INFO:0@0.233] global onnx_importer.cpp:1036 handleNode     Input[3] = 'ConstantOfShape_680'
  [ INFO:0@0.233] global onnx_importer.cpp:1036 handleNode     Input[4] = 'ConstantOfShape_682'
  [ INFO:0@0.233] global onnx_importer.cpp:1040 handleNode     Output[0] = 'NonMaxSuppression_683'

• Test: perf_net.cpp -> DNNTestNetwork, SSD

MobileNet_SSD (opset 10, 12)

• resolved

• Source: https://github.com/onnx/models/tree/main/validated/vision/object_detection_segmentation/ssd-mobilenetv1/model

• Failure:

  opset-10

  [ INFO:0@0.126] global onnx_importer.cpp:803 populateNet DNN/ONNX: loading ONNX v5 model produced by 'tf2onnx':1.7.0. Number of nodes = 350, initializers = 340, inputs = 1, outputs = 4
  [ INFO:0@0.126] global onnx_importer.cpp:696 parseOperatorSet DNN/ONNX: ONNX opset version = 10
  OpenCV(5.0.0-pre) Error: Unsupported format or combination of formats (Unsupported data type: BOOL) in getMatFromTensor, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/onnx/onnx_graph_simplifier.cpp, line 1754
  /Users/wanli/Desktop/OpenCV_China/opencv/modules/ts/src/ts_perf.cpp:1974: Failure

  opset-12

  [ INFO:0@0.159] global onnx_importer.cpp:803 populateNet DNN/ONNX: loading ONNX v7 model produced by 'tf2onnx':1.9.1. Number of nodes = 361, initializers = 346, inputs = 1, outputs = 4
  [ INFO:0@0.159] global onnx_importer.cpp:696 parseOperatorSet DNN/ONNX: ONNX opset version = 12
  [ INFO:0@0.268] global onnx_importer.cpp:974 handleNode DNN/ONNX: processing node with 1 inputs and 1 outputs: [Cast]:(onnx_node!ToFloat) from domain='ai.onnx'
  [ INFO:0@0.269] global onnx_importer.cpp:974 handleNode DNN/ONNX: processing node with 1 inputs and 1 outputs: [Shape]:(onnx_node!Preprocessor/map/Shape) from domain='ai.onnx'
  [ INFO:0@0.270] global onnx_importer.cpp:1036 handleNode     Input[0] = 'ToFloat:0'
  [ INFO:0@0.270] global onnx_importer.cpp:1040 handleNode     Output[0] = 'Preprocessor/map/Shape:0'
  [ERROR:0@0.269] global onnx_importer.cpp:2426 parseShape DNN/ONNX(Shape): dynamic 'zero' shapes are not supported, input ToFloat:0 [ 0 0 0 3 ]
  OpenCV(5.0.0-pre) Error: Assertion failed (!isDynamicShape) in parseShape, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/onnx/onnx_importer.cpp, line 2427
  [ERROR:0@0.270] global onnx_importer.cpp:1032 handleNode DNN/ONNX: ERROR during processing node with 1 inputs and 1 outputs: [Shape]:(onnx_node!Preprocessor/map/Shape) from domain='ai.onnx'
  OpenCV(5.0.0-pre) Error: Unspecified error (> Node [Shape@ai.onnx]:(onnx_node!Preprocessor/map/Shape) parse error: OpenCV(5.0.0-pre) /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/onnx/onnx_importer.cpp:2427: error: (-215:Assertion failed) !isDynamicShape in function 'parseShape'
  > ) in handleNode, file /Users/wanli/Desktop/OpenCV_China/opencv/modules/dnn/src/onnx/onnx_importer.cpp, line 1051

• Test: perf_net.cpp -> DNNTestNetwork, MobileNet_SSD_v1_ONNX

OpenCV Face Detector

• resolved
• Use YuNet instead.
• related PR:
  • Added FaceDetectorYN to OpenCV JS API #25430
  • Change opencv_face_detector related tests and samples from caffe to onnx  #25463
  • Replace opencv_face_detector with YuNet in js_face_recognition sample #25468

GOTURN

• resolved
• The GOTURN model will take 388 MB of traffic for each download if converted to onnx. If the user wants to use the tracking method, we can recommend they use Vit or dasimRPN, which have better performance.
• use VIT Tracker instead
• related PR:
  • Remove goturn caffe model #25503

OpenPose (mpi, coco)

• resolved

• Source: Not Found

• Test: perf_net.cpp -> DNNTestNetwork, OpenPose_pose_mpi_faster_4_stages

• Related PR: Updated openpose sample to run using onnx #25593

fcn8s-heavy-pascal

• resolved

• use fcn-resnet50-12.onnx replace

• need to modify postprocess of the SegmentationModel class

• Test: test_model.cpp -> Test_Model, Segmentation

• Related PR: change fcn8s-heavy-pascal tests from caffe to onnx #25435

faster_rcnn (zf, vgg16)

• ignored
• use FasterRCNN-12.onnx replace
• May be removed the tests in test_int8_layer.cpp

R-FCN

• resolved
• Source: Not Found
• Test: test_model.cpp -> Test_Model, DetectionOutput
• Propose using TensorFlow model. Model link: http://download.tensorflow.org/models/object_detection/rfcn_resnet101_coco_2018_01_28.tar.gz
• Failure
  1. The input name of some layers will be ^name, then cause error in line 267 in graph_simplifier.cpp
  2. Assert layer doesn't support. Because Mat doesn't support String data type.
     

Classic models

• resolved
• alexnet, densenet, googlenet, squeezenet, resnet-50
• all can be found and infered from onnx model zoo
• Related PR: Change classic caffe models related tests and samples to onnx format #25548

layer_convolution and other many single layer tests

• resolved
• need to be converted manually
• Test: test_layer.cpp -> Layer_Test_Convolution_DLDT, Accuracy......
• Related PR: Convert single layer caffe tests to onnx format #25581

hed_pretrained_bsds

• resolved
• https://github.com/s9xie/hed
• sample/dnn/edge_detection.py
• Solutions:
  1. Can be converted to onnx with fp16 (not the best choice, model about 56MB).
  2. I propose to replace with dexined which is the top accuracy in https://paperswithcode.com/sota/edge-detection-on-mdbd , since hed model was released about 10 years ago.
     • Model can be from https://github.com/axinc-ai/ailia-models/tree/master/line_segment_detection/dexined which is MIT License
     • Model link: https://storage.googleapis.com/ailia-models/dexined/model.onnx
• Related PR: Improved edge detection sample #25515

digit_lenet

• resolved
• Can be replaced with https://github.com/ONNC/onnc-tutorial/blob/master/models/lenet/lenet.onnx
• samples/dnn/digitals_lenet.cpp

Colorization

• resolved
• https://github.com/richzhang/colorization
• use model from axinc-ai/ailia-models?
• Source: https://storage.googleapis.com/ailia-models/colorization/colorizer.onnx
• samples/dnn/colorization.py
• Related PR: Replaced caffe model with onnx for colorization sample #25433

FlowNet

• resolved
• replace with RAFT in OpenCV Zoo
• samples/dnn/optical_flow.py

wechat QR code

• resolved
• will be updated by WeChat qBar team

wechat super resolution

• resolved
• will be removed or updated by WeChat qBar team

TextBox & dictnet_vgg

• resolved
• remove / convert to onnx / use other model?
• opencv_contrib/modules/text

3D triplet

• Ignored
• convert to onnx / use other model?
• seems this module don't use OpenCV Caffe importer. Only use official Caffe libary.
• opencv_contrib/modules/cnn_3dobj