versatran01 / galt Goto Github PK

A new training process:

Given n labeled images (raw, neg, pos), divided them in to train, calib, valid.

1. Fit an image pipeline based on train data, and fit a classifier using transformed train data with cross validation
2. Apply the fitted pipeline and classifier on to the calib data and get binary images, these binary images will be processed by a specialized BlobAnalyzer to produce blobs, also the corresponding positive labels will be processed by a basic BlobAnalyzer to produce ground truth blobs. These blobs can be used to calibrate the specialized BlobAnalyzer
3. Apply the fitted pipeline and classifier and the analyzer on to the valid data and compare the final binary image with its corresponding positive label.

Just in case anyone faces this, seems this is a known (and open) bug.

Reproduction: I discovered this problem with stale clock values during reprojection of pointcloud. An easy way to reproduce this is to use rosparam set /use_sim_time true after running roscore, then play back bag (compressed) file with --clock option, and then echo the /clock topic. You will notice the clock count pause randomly after a few seconds.

http://answers.ros.org/question/78676/playback-of-large-compressed-bag-with-clock/?answer=151591
http://answers.ros.org/question/63410/rosbag-not-publishing-clock/?answer=151586

ONGOING TEST: Am about to test on an uncompressed bag file. Will update ticket when I study behavior on it.

the current implementation of bbox_intersect seems to be doing the wrong thing.
Need to investigate.

current one sucks

Eliminate as many prints as possible, except those for explicit report.

Make sure to label all possible apples for self-calibration

• local maxima

http://scikit-image.org/docs/dev/auto_examples/plot_peak_local_max.htm

Add related files

Coverage should be 100%