LEARNING AND INFERENCE IN GRAPHICAL MODELS Chapter 10: Third Working Phase Dr. Martin Lauer University of Freiburg Machine Learning Lab Karlsruhe Institute of Technology Institute of Measurement and Control Systems Learning and Inference in Graphical Models. Chapter 10 p. 1/8
Look back to working phase 1 (MNIST) Accuraccy of approaches on benchmark data: group A (Martin, Cem): 78.47% group B (Simon, Daniel): 51.09% group C (Aaron, Oliver): 41.28% group D (Ramin, Parastou):? Confusion matrix and details on our webpage Learning and Inference in Graphical Models. Chapter 10 p. 2/8
Look back to working phase 1 (MNIST) Accuraccy of approaches on benchmark data: group A (Martin, Cem): 78.47% group B (Simon, Daniel): 51.09% group C (Aaron, Oliver): 41.28% group D (Ramin, Parastou):? Confusion matrix and details on our webpage Do you remember the test errors presented? Learning and Inference in Graphical Models. Chapter 10 p. 2/8
Working phase 3 goal: detection and classification of traffic participants in camera images. KITTI-Objectrecognition-dataset http: //www.cvlibs.net/datasets/kitti/eval_object.php 7481 RGB training images containing approx. 40,000 relevant objects, manually labeled 7518 RGB benchmark images with an unknown number of objects object classes: Car (28742), Van (2914), Truck (1094), Tram (511), Cyclist (1627), Pedestrian (4487), Person sitting (222), Misc (973) Learning and Inference in Graphical Models. Chapter 10 p. 3/8
Benchmarking KITTI offers a possibility to evaluate and compare your final results on the benchmark image set. you apply your final approach on the benchmark images and write one label file with the detected objects KITTI provides accuracy measures for your detections and classifications comparison with other methods provided on the KITTI webpage the role of the DontCare objects: these show potentially relevant objects, however, they are not used for the performance evaluation (e.g. object too small, too much occluded, etc.) non-maxima suppression: for each object only one bounding box should be returned. This function will be provided by us (somewhen during the next weeks) Learning and Inference in Graphical Models. Chapter 10 p. 4/8
What you should do develop an approach to detect and classify traffic participants in RGB images based on the training images provided by KITTI potentially using other data sources (cf. KITTI webpage) potentially using image distortion and preprocessing techniques (color transformation, filtering, feature calculation/selection) using one or more approaches of deep learning architectures Learning and Inference in Graphical Models. Chapter 10 p. 5/8
What you should do develop an approach to detect and classify traffic participants in RGB images based on the training images provided by KITTI potentially using other data sources (cf. KITTI webpage) potentially using image distortion and preprocessing techniques (color transformation, filtering, feature calculation/selection) using one or more approaches of deep learning architectures until 9. Feb. 2014, 24:00 apply your best approach on the benchmark images and create one label file for each benchmark image zip the label files (using zip or tar/gnuzip=tgz) send the zip archive to me by email (martin.lauer@kit.edu) Learning and Inference in Graphical Models. Chapter 10 p. 5/8
What you should do develop an approach to detect and classify traffic participants in RGB images based on the training images provided by KITTI potentially using other data sources (cf. KITTI webpage) potentially using image distortion and preprocessing techniques (color transformation, filtering, feature calculation/selection) using one or more approaches of deep learning architectures until 9. Feb. 2014, 24:00 apply your best approach on the benchmark images and create one label file for each benchmark image zip the label files (using zip or tar/gnuzip=tgz) send the zip archive to me by email (martin.lauer@kit.edu) until 12. Feb. 2014, 12:30 prepare a presentation (10 minutes+questions) on your approach present your slides in the final lesson on 12. Feb. 2014 Learning and Inference in Graphical Models. Chapter 10 p. 5/8
Some technical notes where do you get the data? KITTI webpage www.cvlibs.net/datasets/kitti/eval_object.php download left images of object data set (12GB) download training labels of object data (5MB) if you like, download the object development kit (contains a readme file that explains the label file format) Learning and Inference in Graphical Models. Chapter 10 p. 6/8
Some technical notes where do you get the data? KITTI webpage www.cvlibs.net/datasets/kitti/eval_object.php download left images of object data set (12GB) download training labels of object data (5MB) if you like, download the object development kit (contains a readme file that explains the label file format) tools for reading writing images and label files in Lua/Torch and display results is provided on our webpage (kitti tools.lua) read_image loads a KITTI image with given file path and image number read_labels loads a KITTI label file and provides a table of objects write_labels writes a KITTI label file with a given set ob objects add_labels_to_images creates an image that displays the detected objects with bounding boxes Learning and Inference in Graphical Models. Chapter 10 p. 6/8
Some technical notes each object is represented in Lua/Torch as a table with attributes: type: string that indicates the class label ( Car, Pedestrian,...) x1, y1, x2, y2: the pixel coordinates of the left, upper, right, lower bound of the bounding box around the object score: a real number the indicates the confidence of your classification. The more confident the classifier is the larger the number should be (is used to calculate precision/recall curves) truncated: a number between 0 and 1 that indicates up to which extend the object is truncated at the image boundaries (0=non-truncated, 1=largely truncated) occluded: an integer between 0 and 3 that indicates whether the object is non-occluded (0), partly occluded (1), largely occluded (2) or the amount of occlusion is unknown (3) Learning and Inference in Graphical Models. Chapter 10 p. 7/8
Grouping group β: Martin, Ramin group γ: Oliver, Simon, Aaron group δ: Daniel, Cem Learning and Inference in Graphical Models. Chapter 10 p. 8/8