bit-bots / bitbots_vision Goto Github PK

In idle mode the rate at which images are processed should be drastically reduced

colorpicker, hsv colorpicker and colorspace tool could be merged into one tool

We could use the Camera position to estimate a horizon in the picture without inspecting it. This could increase performance.

evaluate the vision params for optimal performance on the jetson.

For localization it would be nice to also have a list of points, on which no lines were detected.

Currently, only the top candidate gets added to the ball candidates message, but the message is designed to contain all candidates.

To save some resources while the RoboCup world championship 2019, the parameter line_detector_linepoints_range has been set to 0.
This parameter determines the number of created line points.

Please reset this afterward to default 200.

Instead of handling, whether the vision is use for simulation or not, inside the vision, the parameters should be overwritten by an additional simulation parameter file in the launch file.

In case the config value for the ball detector does not match any of the detectors what should be the behavior. Add a dummy detector?

Please notify the user to rebuild the vision node, if the colorspaces or models folder has changed.

This is referring to #46.

As in #43 there should be a proper way to disable or reduce the number of non line points.

There should be a parameter set for the „Feldraum“ that can easily be used, for example with feldraum:=true.

Add arguments like basler:=true and logitech:=true to the vision launch file

• driver ✓
• launch file
• adapt camera calibration (see bit-bots/wolves_image_provider#2)
• make everything work

Currently, print is used in the live_classifier and live_fcnn_03 files. The debug printer should be used instead.

Currently, the vision debug printer just reimplements the actual ROS named loggers. You can provide a logger name like this: rospy.logdebug('message', logger_name='field_boundary_detector'). All logged messages will be shown and can be filtered int the rqt log console and the logger leven in the actual screen output can be controlled by using the rqt logger level plugin or by calling the set_logger_level service. The logger name in the above example would be /rosout/<node name>/field_boundary_detector.
Using these loggers would also reduce external dependencies of the modules.

Create separate classes for field boundary search methods. Therefore inherit from "basic" field boundary detector.

Currently, we are using the head tilt angle to determine whether or not we will use the reversed field boundary search method. It would be nicer to get this angle from the transform from the base_footprint to the camera_optical_frame. In this case we are considering the whole kinematic chain and drop the config parameter because the real horizon always occurs at 90 degrees.

This Repo needs a slight refactoring to clean up the code, add comments and merge the bitbots_vision directory with the bitbots_vision_common directory.

Do not cache any data of field boundary detector while dynamic color space is running, because the color space of the color detector could have changed in the meantime.

Publish lines as a binary image mask.

This issue is related to #64.
As the message format is defind to for multiple balls, the topic should be named "balls_in_image".

The FCNN has to be retrained on a new image set with and without the ball from the robot's point of view.

FCNN handler: add param for publishing debug image like in the color detector

Also add this parameter to the launch file.

The line detector throws a numpy exception if no line points are detected.

We should only reconfigure changed parameters in the dynamic reconfigure callback.
Some checks have recently been added in pull request #59.

I just stumbled upon resource_retriever and it looks pretty useful.

Can you utilize it so that for example colorspaces dont have to be specified as an absolute path?

Referring to a code comment, we should find a better name for the function "_obstacle_detector_distance".

The FCNN has to be evaluated with the new camera. Possibly, a blur has to be added. This issue is related to #23, #22 and #20.

The horizon detector needs to work more stable and a bit faster. The most dangerous Situation are fields next to each other.

currently, we are not using any information saved from the last image. When we do, we need to be able to reset it all.

add description for vision tools

Originally posted by @jangut in #34

We should evaluate in which cases we want to use the runtime evaluator in the future. The current solution with the evaluator removed via code comments seems not like a permanent solution. Maybe we enhance it further (with activation via the launch parameter/config) or we use common profiling tools like profilehooks.

The PixelListColorDetector should not be dynamic. But another new detector inheriting from it should.

• via FCNN (@johagge)
• via conventional methods (@hendrikbrandt)

Please rename field boundary finder to detector in config and cfg files

Config files need more comments describing every parameters effect.

We should evaluate the ROS log levels. This is related to #18.