CVS-160560 nanobindings for keypoint detection and segmentation #258
Conversation
| nb::class_<InstanceSegmentationResult, ResultBase>(m, "InstanceSegmentationResult") | ||
| .def(nb::init<int64_t, std::shared_ptr<MetaData>>(), nb::arg("frameId") = -1, nb::arg("metaData") = nullptr) | ||
| .def_prop_ro( | ||
| "feature_vector", |
There was a problem hiding this comment.
Iseg result also contains a saliency map as a vector of cv::Mat. Perhaps, it's not that easy to expose but there is a chance a vector of nb::ndarray<uint8_t... would work here
| .def_prop_ro( | ||
| "soft_prediction", | ||
| [](ResultBase& r) { | ||
| ImageResultWithSoftPrediction ir = r.asRef<ImageResultWithSoftPrediction>(); |
There was a problem hiding this comment.
Would that work if the result entity is actually an ImageResult?
In sseg wrapper now we have the following logic:
if (return_soft_prediction) {
return a ptr to ImageResultWithSoftPrediction
}
else
return a ptr to ImageResult
infer() method's return type is a ptr to ImageResult. I think that kind of design looks weird an overall inconvenient, because we always need to check if type conversion to ImageResultWithSoftPrediction is successful.
Related refactoring can be done in subsequent PR, here we just need to be sure that both the cases return_soft_prediction==true/false are handled correctly
|
Closed as outdated |
What does this PR do?
Ready for python replacement, this will have the same structure results as python version:
Still needs work: