Florian Jaton - The Constitution of Algorithms

List of Figures

The CSF main building. On the left and right sides of the central patio, lines of offices and seminar rooms. In the center of the image, in air-conditioned rooms with glazed windows, three server farms store local programs, experiments, and databases. On the top floor, illuminated, one can discern the entrance to the faculty cafeteria.

The Labs hall. On the left, behind closed doors, the Labs cafeteria and seminar room. On the right, seven offices most of the time occupied by two researchers.

Inside one of the Labs offices. Two researchers were generally facing each other, though they were behind one to three large monitors.

Schematic of the pixel organization of a digital photograph as enabled by industrially produced and standardized CCDs and CMOSs. The schematic on the right is an imaginary zoom of the digital photograph on the left. Every pixel is identified by its location within a coordinate system (x/y). Moreover, assuming the image on the left is a color image, each pixel is described by three complementary values, commonly referred to as a red, green, and blue (RGB) color scheme. As most standard computers now express RGB values as eight-bit memory addresses (e.g., one byte), these triplets can vary from zero to 255 or, in hexadecimal writing, from 00 to FF.

Example of an academic paper published by an industrial research team. This paper dealing with deep neural networks for image recognition won the best paper award of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Though copyrighted by the Institute of Electrical and Electronics Engineers (IEEE) (the official editor of the conferences proceedings), its content is freely available in the arXiv.org repository. Source: He et al., 2016. Reproduced with permission from IEEE.

Excerpt from one of my logbooks and its translation into a .txt file. On the left, notes taken during a Lab meeting on November 16, 2014. On the right, the translation of these notes into a .txt file. The name of the file starts with l-meeting, thus indicating it refers to a Lab meeting. The second section, 141106, refers to the date of the logbook entry. The third section, nk, refers to the initials of the collaborator the note concerns. The fourth section, deep-learning-on-manuscripts, refers to the title of the presentation. The fifth and last section (l42738) indicates the location of the original note, here in logbook number 4, from page 27 to page 38.

Example of a small Python script used to browse the content of the .txt files. This script, working as a small computer program, makes the computer list the names of the .txt files whose content include the keywords ground truth and NK in a new document named 0_list-entries.

Schematic of precision and recall measures on . In this hypothetical example, (grey background) detected thirty targets (true positives) but missed eighteen of them (false negatives). This performance means that has a recall score of 0.62. The algorithm also detected twelve elements that are not targets (false positives), and this makes it have a precision score of 0.71. From this point, other algorithms intended to detect the same targets can be tested on the same ground truth and may have better or worse precision and recall scores than .

An exemplary comparison table among high-level face-detection algorithms. Two ground truths are used for this comparison table from Carnegie Mellon University (CMU) and the Massachusetts Institute of Technology (MIT). On the left, a list of algorithms named according to the papers in which they were proposed. In this table, the Percentage of Correct Detection (CD) indicates the recall values and the Number of False Positives (FP) suggests the precision values. Source: Hjelms and Low (2001, 262). Reproduced with permission from Elsevier.

Samples from Liu et al.s dataset. Pictures contain one centered and contrastive element. Source: Microsoft Research Asia (MSRA) public dataset, Liu et al. (2007).

Performance evaluations on Liu et al.s ground truth. On the left, a visual comparison among three different saliency-detection algorithms according to the ground truth. On the right, histograms that summarize the statistical performances of the three algorithms. In these histograms, the ground truth corresponds to the y axis, the best possible saliency-detection performance that enables the evaluation. Source: Liu et al. (2007, 7). Reproduced with permission from IEEE.

Image (a) is an unlabeled image of Liu et al.s ground truth; image (b) is the result of Wang & Lis saliency-detection algorithm; image (c) is the imaginary result of some other saliency-detection algorithm on (a); and image (d) is the bounding-box target as provided by Liu et al.s ground truth. Even though (b) is more accurate than (c), it will obtain a lower statistical evaluation if compared to (d). This is why Wang & Li propose (e), a binary target that matches the contours of the already defined salient object. Source: Wang and Li (2008, 968). Reproduced with permission from IEEE.

2013 comparison table between different saliency-detection algorithms. The number of competing algorithms has increased since 2007. Here, three ground truths are used for performance evaluations: ASD (Achanta et al. 2009), SED (Alpert et al. 2007), and SOD (Movahedi and Elder 2010). Below the figure, a table compares the execution time of each implemented algorithm. Source: Jiang et al. (2013, 1672). Reproduced with permission from IEEE.

Screen captures of the web application designed by the Group for its crowdsourcing task. On the left, the application when ran by a web browser. Once workers created a username, they could start the experiment and draw rectangles. When workers clicked on Next Image button, the coordinates of the rectangles were stored in .txt files on the Labs server. On the right, one excerpt of one of the seven scripts required to realize such interactive labels and data storage.