-
-
Notifications
You must be signed in to change notification settings - Fork 3.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Tagging objects for semantic segmentation suppresses render occlusion #4
Comments
For now we disable custom depth pass for each object if semantic segmentation is not needed. However, it is still active in the project settings by default. If necessary, can be disabled (in the release binary) by adding the file "CarlaUE4/Saved/Config/LinuxNoEditor/Engine.ini" containing the following [/Script/Engine.RendererSettings]
r.CustomDepth=1 |
This also makes to make an extra pass for the custom depth render, thus is sucking performance. The solution EPIC proposes is to make a material property to be applied in all the materials of all the objects in the world and then set the values for the stencil with the id of the object type for semantic segmentation, that will be used later in the post-fx shader...not an easy way to do it though |
I guess there are no update for this issue, right? |
@germanros1987 not yet, this issue is a particular hard one, considering the other option is to change the core of the engine instead of doing the material trick, we are waiting for another solution from the Epic team they are digging into the problem btw. |
@bernatx Is this still an active problem? |
The problem cannot be seen in the cameras, the thing is that UE4 stops doing occlusion culling for meshes with the stencil buffer enabled (used to encode semseg). The resulting image is the same, but you're rendering more meshes than necessary, so it's just a performance issue. I have no idea if it's still happening but I suppose it is, since the main use for the stencil buffer is precisely to render occluded objects, probably the only solution is to use a different mechanism to render semantic segmentation. |
It would be helpful if there is a tracking/bug number from Epic for reference. Thanks! |
Of course I was overexcited and still, we need to push harder on this, other companies in UDN are requesting the same thing, to control whether or not it is subject to Occlusion Culling on an actor-by-actor basis, not limited to Custom Depth.
|
Merge in VIAD/carla from feature/merge_assets to dev_0.9.13 * commit '75d908dc75870ab5dceac7d7aaac7b79bc143eb5': Change scale to 0.85 * Audi tt : Set car settings (Torque, Engine, Gears, mass, ...) as they were in 0.9.9 (master_aw) * BaseVehiclePawn : Changed the input nodes with the correct ones Some corrections for the Fanatec switches setup Rendering settings from master_aw (except for forward shading which breaks Unreal) Town 03 : replaced american traffic lights with european looking ones Town 04 : removed 3 traffic lights on the highway A little cleanup and setup Remove the double Audi tt entry in VehicleFactory vehicles list after reanming/replacing the Audi_tt_Carla0913_viaduct one back to Audi_tt Remove original audi tt asset Replaces and moved all "american" traffic lights with a more european looking one. Hid some traffic lights on a "highway" part of the road: Z value changed: AmericanLights_Town04_02 : 0 -> -900 AmericanLights_Town04_03 : 0 -> -900 AmericanLightsTown04_17 : 0 -> -900 Carla 0.9.13 Original version of Town03 and Town04. New AudiTT asset which is the result of a "merge" between the current AudiTT BP and the Audi TT BP from Carla 0.9.13 Update Vehicles list: * Added a reference to BP_AudiTT_Carla0913_Viaduct (it is a merge bewteen BP_Audi_TT from dev 0.9.12 and the modifications by Carla 0.9.13) * Added the 2 cars from Carla 0.9.13 (Frod Crown and Volkswagen T2) Function ApplyColor : ported the Carla 0.9.13 modifications regarding change of doors colors.
When using CARLA with Python 3.10, I'm getting a segfault in GetAvailableMaps. The problem disappears when PyList manipulation does not happen with GIL unlocked, as done in this commit. The initial part of crash backtrace (from GDB) is below: Program terminated with signal SIGSEGV, Segmentation fault. warning: Section `.reg-xstate/49253' in core file too small. #0 _PyInterpreterState_GET () at ./Include/internal/pycore_pystate.h:117 117 return tstate->interp; [Current thread is 1 (Thread 0x7fe6fe48f740 (LWP 49253))] (gdb) bt #0 _PyInterpreterState_GET () at ./Include/internal/pycore_pystate.h:117 matejm42#1 get_list_state () at Objects/listobject.c:26 carla-simulator#2 PyList_New (size=0) at Objects/listobject.c:159 carla-simulator#3 0x00007fe6fdc0dab0 in boost::python::detail::list_base::list_base() () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 carla-simulator#4 0x00007fe6ef9ecfc4 in boost::python::list::list (this=0x7ffd8a8aae28) at include/boost/python/list.hpp:61 carla-simulator#5 GetAvailableMaps (self=...) at source/libcarla/Client.cpp:26 carla-simulator#6 0x00007fe6efb6a8fe in boost::python::detail::invoke<boost::python::to_python_value<boost::python::list const&>, boost::python::list (*)(carla::client::Client const&), boost::python::arg_from_python<carla::client::Client const&> > (rc=..., f=<optimized out>, ac0=...) at include/boost/python/detail/invoke.hpp:73 carla-simulator#7 boost::python::detail::caller_arity<1u>::impl<boost::python::list (*)(carla::client::Client const&), boost::python::default_call_policies, boost::mpl::vector2<boost::python::list, carla::client::Client const&> >::operator() (args_=<optimized out>, this=<optimized out>) at include/boost/python/detail/caller.hpp:233 carla-simulator#8 boost::python::objects::caller_py_function_impl<boost::python::detail::caller<boost::python::list (*)(carla::client::Client const&), boost::python::default_call_policies, boost::mpl::vector2<boost::python::list, carla::client::Client const&> > >::operator() ( this=<optimized out>, args=<optimized out>, kw=<optimized out>) at include/boost/python/object/py_function.hpp:38 carla-simulator#9 0x00007fe6fdc1b4dd in boost::python::objects::function::call(_object*, _object*) const () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 carla-simulator#10 0x00007fe6fdc1b6a8 in boost::detail::function::void_function_ref_invoker0<boost::python::objects::(anonymous namespace)::bind_return, void>::invoke(boost::detail::function::function_buffer&) () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 ...
When using CARLA with Python 3.10, I'm getting a segfault in GetAvailableMaps. The problem disappears when PyList manipulation does not happen with GIL unlocked, as done in this commit. The initial part of crash backtrace (from GDB) is below: Program terminated with signal SIGSEGV, Segmentation fault. warning: Section `.reg-xstate/49253' in core file too small. #0 _PyInterpreterState_GET () at ./Include/internal/pycore_pystate.h:117 117 return tstate->interp; [Current thread is 1 (Thread 0x7fe6fe48f740 (LWP 49253))] (gdb) bt #0 _PyInterpreterState_GET () at ./Include/internal/pycore_pystate.h:117 matejm42#1 get_list_state () at Objects/listobject.c:26 carla-simulator#2 PyList_New (size=0) at Objects/listobject.c:159 carla-simulator#3 0x00007fe6fdc0dab0 in boost::python::detail::list_base::list_base() () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 carla-simulator#4 0x00007fe6ef9ecfc4 in boost::python::list::list (this=0x7ffd8a8aae28) at include/boost/python/list.hpp:61 carla-simulator#5 GetAvailableMaps (self=...) at source/libcarla/Client.cpp:26 carla-simulator#6 0x00007fe6efb6a8fe in boost::python::detail::invoke<boost::python::to_python_value<boost::python::list const&>, boost::python::list (*)(carla::client::Client const&), boost::python::arg_from_python<carla::client::Client const&> > (rc=..., f=<optimized out>, ac0=...) at include/boost/python/detail/invoke.hpp:73 carla-simulator#7 boost::python::detail::caller_arity<1u>::impl<boost::python::list (*)(carla::client::Client const&), boost::python::default_call_policies, boost::mpl::vector2<boost::python::list, carla::client::Client const&> >::operator() (args_=<optimized out>, this=<optimized out>) at include/boost/python/detail/caller.hpp:233 carla-simulator#8 boost::python::objects::caller_py_function_impl<boost::python::detail::caller<boost::python::list (*)(carla::client::Client const&), boost::python::default_call_policies, boost::mpl::vector2<boost::python::list, carla::client::Client const&> > >::operator() ( this=<optimized out>, args=<optimized out>, kw=<optimized out>) at include/boost/python/object/py_function.hpp:38 carla-simulator#9 0x00007fe6fdc1b4dd in boost::python::objects::function::call(_object*, _object*) const () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 carla-simulator#10 0x00007fe6fdc1b6a8 in boost::detail::function::void_function_ref_invoker0<boost::python::objects::(anonymous namespace)::bind_return, void>::invoke(boost::detail::function::function_buffer&) () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 ...
When I run generate_traffic.py under Python 3.10, I get a segfault at line: loc = world.get_random_location_from_navigation() The backtrace from gdb looks like this: #0 0x00007f04552ad7e7 in new_threadstate () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/libpython3.10.so.1.0 matejm42#1 0x00007f04552adaa1 in PyGILState_Ensure () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/libpython3.10.so.1.0 carla-simulator#2 0x00007f040afd4f32 in std::_Function_handler<void (carla::client::WorldSnapshot), MakeCallback(boost::python::api::object)::{lambda(auto:1)matejm42#1}>::_M_invoke(std::_Any_data const&, carla::client::WorldSnapshot&&) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#3 0x00007f040b1d4ab1 in carla::client::detail::CallbackList<carla::client::WorldSnapshot>::Call(carla::client::WorldSnapshot) const () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#4 0x00007f040b1d424a in std::_Function_handler<void (carla::Buffer), carla::client::detail::Episode::Listen()::{lambda(auto:1)matejm42#1}>::_M_invoke(std::_Any_data const&, carla::Buffer&&) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#5 0x00007f040b23fc41 in boost::asio::detail::completion_handler<boost::asio::detail::binder0<carla::streaming::detail::tcp::Client::ReadData()::{lambda()matejm42#1}::operator()() const::{lambda(boost::system::error_code, unsigned long)matejm42#1}::operator()(boost::system::error_code, unsigned long) const::{lambda()matejm42#1}>, boost::asio::io_context::basic_executor_type<std::allocator<void>, 0ul> >::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#6 0x00007f040b24ae85 in boost::asio::detail::strand_service::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#7 0x00007f040b1a94f5 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#8 0x00007f040b199351 in boost::asio::detail::scheduler::run(boost::system::error_code&) [clone .isra.0] () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#9 0x00007f040b1ac1cb in std::thread::_State_impl<std::thread::_Invoker<std::tuple<carla::ThreadPool::AsyncRun(unsigned long)::{lambda()matejm42#1}> > >::_M_run() () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so carla-simulator#10 0x00007f040bce05c3 in execute_native_thread_routine () from /nix/store/2fpmbk0g0ggm9zq89af7phvvvv8dnm7n-gcc-12.3.0-lib/lib/libstdc++.so.6 carla-simulator#11 0x00007f045509fdd4 in start_thread () from /nix/store/1x4ijm9r1a88qk7zcmbbfza324gx1aac-glibc-2.37-8/lib/libc.so.6 carla-simulator#12 0x00007f04551219b0 in clone3 () from /nix/store/1x4ijm9r1a88qk7zcmbbfza324gx1aac-glibc-2.37-8/lib/libc.so.6 It turns out that its caused by releasing GIL for too long. We fix it by releasing the GIL only for the actual libcarla call and constructing Python objects with GIL locked.
* PythonAPI: Fix segfault in GetAvailableMaps When using CARLA with Python 3.10, I'm getting a segfault in GetAvailableMaps. The problem disappears when PyList manipulation does not happen with GIL unlocked, as done in this commit. The initial part of crash backtrace (from GDB) is below: Program terminated with signal SIGSEGV, Segmentation fault. warning: Section `.reg-xstate/49253' in core file too small. #0 _PyInterpreterState_GET () at ./Include/internal/pycore_pystate.h:117 117 return tstate->interp; [Current thread is 1 (Thread 0x7fe6fe48f740 (LWP 49253))] (gdb) bt #0 _PyInterpreterState_GET () at ./Include/internal/pycore_pystate.h:117 #1 get_list_state () at Objects/listobject.c:26 #2 PyList_New (size=0) at Objects/listobject.c:159 #3 0x00007fe6fdc0dab0 in boost::python::detail::list_base::list_base() () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 #4 0x00007fe6ef9ecfc4 in boost::python::list::list (this=0x7ffd8a8aae28) at include/boost/python/list.hpp:61 #5 GetAvailableMaps (self=...) at source/libcarla/Client.cpp:26 #6 0x00007fe6efb6a8fe in boost::python::detail::invoke<boost::python::to_python_value<boost::python::list const&>, boost::python::list (*)(carla::client::Client const&), boost::python::arg_from_python<carla::client::Client const&> > (rc=..., f=<optimized out>, ac0=...) at include/boost/python/detail/invoke.hpp:73 #7 boost::python::detail::caller_arity<1u>::impl<boost::python::list (*)(carla::client::Client const&), boost::python::default_call_policies, boost::mpl::vector2<boost::python::list, carla::client::Client const&> >::operator() (args_=<optimized out>, this=<optimized out>) at include/boost/python/detail/caller.hpp:233 #8 boost::python::objects::caller_py_function_impl<boost::python::detail::caller<boost::python::list (*)(carla::client::Client const&), boost::python::default_call_policies, boost::mpl::vector2<boost::python::list, carla::client::Client const&> > >::operator() ( this=<optimized out>, args=<optimized out>, kw=<optimized out>) at include/boost/python/object/py_function.hpp:38 #9 0x00007fe6fdc1b4dd in boost::python::objects::function::call(_object*, _object*) const () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 #10 0x00007fe6fdc1b6a8 in boost::detail::function::void_function_ref_invoker0<boost::python::objects::(anonymous namespace)::bind_return, void>::invoke(boost::detail::function::function_buffer&) () from /nix/store/c95f3nrkz3sflvycihyw1c8q4nk47p4m-boost-1.79.0/lib/libboost_python310.so.1.79.0 ... * PythonAPI: Fix segfault in get_random_location_from_navigation() When I run generate_traffic.py under Python 3.10, I get a segfault at line: loc = world.get_random_location_from_navigation() The backtrace from gdb looks like this: #0 0x00007f04552ad7e7 in new_threadstate () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/libpython3.10.so.1.0 #1 0x00007f04552adaa1 in PyGILState_Ensure () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/libpython3.10.so.1.0 #2 0x00007f040afd4f32 in std::_Function_handler<void (carla::client::WorldSnapshot), MakeCallback(boost::python::api::object)::{lambda(auto:1)#1}>::_M_invoke(std::_Any_data const&, carla::client::WorldSnapshot&&) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #3 0x00007f040b1d4ab1 in carla::client::detail::CallbackList<carla::client::WorldSnapshot>::Call(carla::client::WorldSnapshot) const () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #4 0x00007f040b1d424a in std::_Function_handler<void (carla::Buffer), carla::client::detail::Episode::Listen()::{lambda(auto:1)#1}>::_M_invoke(std::_Any_data const&, carla::Buffer&&) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #5 0x00007f040b23fc41 in boost::asio::detail::completion_handler<boost::asio::detail::binder0<carla::streaming::detail::tcp::Client::ReadData()::{lambda()#1}::operator()() const::{lambda(boost::system::error_code, unsigned long)#1}::operator()(boost::system::error_code, unsigned long) const::{lambda()#1}>, boost::asio::io_context::basic_executor_type<std::allocator<void>, 0ul> >::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #6 0x00007f040b24ae85 in boost::asio::detail::strand_service::do_complete(void*, boost::asio::detail::scheduler_operation*, boost::system::error_code const&, unsigned long) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #7 0x00007f040b1a94f5 in boost::asio::detail::scheduler::do_run_one(boost::asio::detail::conditionally_enabled_mutex::scoped_lock&, boost::asio::detail::scheduler_thread_info&, boost::system::error_code const&) () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #8 0x00007f040b199351 in boost::asio::detail::scheduler::run(boost::system::error_code&) [clone .isra.0] () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #9 0x00007f040b1ac1cb in std::thread::_State_impl<std::thread::_Invoker<std::tuple<carla::ThreadPool::AsyncRun(unsigned long)::{lambda()#1}> > >::_M_run() () from /nix/store/zqj9irpw63pal9r4671p1gjd9jiw5sid-ros-env/lib/python3.10/site-packages/carla/libcarla.cpython-310-x86_64-linux-gnu.so #10 0x00007f040bce05c3 in execute_native_thread_routine () from /nix/store/2fpmbk0g0ggm9zq89af7phvvvv8dnm7n-gcc-12.3.0-lib/lib/libstdc++.so.6 #11 0x00007f045509fdd4 in start_thread () from /nix/store/1x4ijm9r1a88qk7zcmbbfza324gx1aac-glibc-2.37-8/lib/libc.so.6 #12 0x00007f04551219b0 in clone3 () from /nix/store/1x4ijm9r1a88qk7zcmbbfza324gx1aac-glibc-2.37-8/lib/libc.so.6 It turns out that its caused by releasing GIL for too long. We fix it by releasing the GIL only for the actual libcarla call and constructing Python objects with GIL locked. --------- Co-authored-by: bernat <bernatx@gmail.com>
Currently we tag objects by setting their custom depth stencil value, this way it automatically works for semantic segmentation as well. It turns out that setting this value makes objects being rendered even if they are behind another object. This occurs even if the custom depth pass is disabled at project level.
The text was updated successfully, but these errors were encountered: