Differences

This shows you the differences between two versions of the page.

--- courses:b3b33vir:tutorials:hw2:start [2018/11/09 16:36]
petrito1 [Goals]
+++ courses:b3b33vir:tutorials:hw2:start [2018/11/13 16:50]
petrito1 [Goals]
@@ Line 5: / Line 5: @@
   - Download {{:courses:b3b33vir:tutorials:hw2:hw2.zip|source code package}} and implement missing parts in module ''intrinsics.py'' (see below)
   - Get familiar with pinhole camera model, nonlinear distortion model and respective [[https://docs.opencv.org/3.4.3/d9/d0c/group__calib3d.html|calibration routines from OpenCV]]
-  - Calibrate camera intrinsic parameters (rational model without tangetial distortion)
+  - Calibrate camera intrinsic parameters
     * Implement function
       * ''intrinsics.calibrate''
     * Using functions
       * ''cv2.findChessboardCorners''
-      * ''cv2.cornerSubPix''
+      * ''cv2.cornerSubPix'' (limit search window size ''winSize'' to avoid excessive corrections)
-      * ''cv2.calibrateCamera''
+      * ''cv2.calibrateCamera'' (use the rational model without tangential distorion from the lecture: ''flags=cv2.CALIB_RATIONAL_MODEL + cv2.CALIB_ZERO_TANGENT_DIST'')
   - Convert camera intrinsics parameters (camera matrix and field of view)
     * Implement functions
@@ Line 36: / Line 36: @@
 python multimodal_dataset.py calibrate
 </code>
-Download the multimodal dataset and calibrate intrinsic parameters of its cameras (calls ''intrinsics.calibrate'').
+Download the multimodal dataset [1] and calibrate intrinsic parameters of its cameras (calls ''intrinsics.calibrate'').
 The calibration is saved into the following JSON files:
   * ''data/multimodal_ir_intrinsics.json''
@@ Line 66: / Line 66: @@
 Undistort images using the calibration (calls ''intrinsics.remap'' and ''intrinsics.camera_field_of_view'').
 Instead of specifying the parameters manually, an optimal camera parameters are estimated to ensure all pixels are valid (''%%--alpha 0%%'') or no information is lost (''%%--alpha 1%%''). Any value in-between can also be used. \\
+===== Expected Results =====
+Expected re-projection errors (the 1st output of ''cv2.calibrateCamera'') for the multimodal dataset are approximatelly:
+  * Left camera:  0.37 px (0.67 px without sub-pixel refinement via ''cv2.cornerSubPix'')
+  * Right camera: 0.33 px (0.68 px)
+  * IR camera: 0.78 px (1.27 px)
+Original and undistorted images from the left camera should look similarly to these:
+| {{ :courses:b3b33vir:tutorials:hw2:2010_10_19_14_05_46_bb_left.png?direct&400 |}} | {{ :courses:b3b33vir:tutorials:hw2:2010_10_19_14_05_46_bb_left_45_480x640.png?direct&400 |}} | {{ :courses:b3b33vir:tutorials:hw2:2010_10_19_14_05_46_bb_left_46_480x640.png?direct&400 |}} |
+| Original | Undistorted, ''%%--alpha 0%%'' | Undistorted, ''%%--alpha 1%%'' |
+===== References =====
+[1] Barrera F., Lumbreras F., Sappa A. Multimodal Stereo Vision System: 3D Data Extraction and Algorithm Evaluation. In //IEEE Journal of Selected Topics in Signal Processing//, Vol. 6, No. 5, September 2012, pp. 437--446.