Medium/2
The goal of this contest is to improve a strong classifier for predicting the classes of the CIFAR-10 dataset.
This strong classifier was produced by the experiment Medium/2/base with 100 iterations of Boosting using 48 heuristics implemented by the consortium and the winner of the previous contest. It uses features from 25 heuristics, and reaches a test error of 47.78%.
Each heuristic participating in the contest will be evaluated by running 100 additional iterations of Boosting using only this new heuristic. Performance will be estimated as the gain in test error between the resulting strong classifier and the original one.
At the end of the contest (will be announced later), the heuristic leading to the lowest test error rate will be declared the winner of the contest. It will be used to train a new strong classifier, together with the consortium heuristics. This classifier will then be used to start a new contest. If no participating heuristic improves the test error, the duration of the contest will be extended.
Anyone can participate to this contest. The heuristics must be public, but their source code will not be displayed before the date chosen by their author, which can be as far as four weeks in the future. Heuristics written by the consortium members cannot compete against the ones from the contributors, but their performance appears in a separate ranking.
| - | Entries (1) |
| RANK | DELTA | HEURISTIC | DESCRIPTION | EXPERIMENT | RESULTS | JOIN DATE |
|---|---|---|---|---|---|---|
| 1 | -1.89% | wenqi/hue/1 | Medium/2/wenqi/hue | 44.62% / 45.89% | Oct. 25, 2011, 11:25 a.m. |
| - | Consortium entries (48) |
| RANK | DELTA | HEURISTIC | DESCRIPTION | EXPERIMENT | RESULTS | JOIN DATE |
|---|---|---|---|---|---|---|
| 1 | -4.95% | leonidas/rgb_hogs/2 | With color | Medium/2/leonidas/rgb_hogs/2 | 41.73% / 42.83% | Dec. 27, 2011, 8:45 p.m. |
| 2 | -4.22% | leonidas/rgb_hogs/1 | With color | Medium/2/leonidas/rgb_hogs | 42.03% / 43.56% | Oct. 25, 2011, 11:24 a.m. |
| 3 | -4.19% | leonidas/corner_scratches256/2 | Medium/2/leonidas/corner_scratches256/2 | 42.45% / 43.59% | Oct. 25, 2011, 11:24 a.m. | |
| 4 | -4.13% | cdubout/v1plus/1 | Medium/2/cdubout/v1plus | 41.87% / 43.65% | Oct. 25, 2011, 11:23 a.m. | |
| 5 | -4.00% | leonidas/corner_scratches512/2 | Medium/2/leonidas/corner_scratches512/2 | 42.33% / 43.78% | Oct. 25, 2011, 11:24 a.m. | |
| 6 | -3.95% | cdubout/mb_ilbp/4 | Pyramid of histograms of multi-block improved local binary patterns. Inspired from ftarsett/ilbp and created with the help of Cosmin Atanasoei. | Medium/2/cdubout/mb_ilbp/4 | 42.71% / 43.83% | Oct. 25, 2011, 11:23 a.m. |
| 7 | -3.76% | leonidas/dt_hogs6/1 | Histograms of Oriented Gradients (Dalal and Triggs) | Medium/2/leonidas/dt_hogs6 | 42.60% / 44.02% | Oct. 25, 2011, 11:24 a.m. |
| 8 | -3.66% | leonidas/corner_scratches1600/2 | Medium/2/leonidas/corner_scratches1600/2 | 42.33% / 44.12% | Oct. 25, 2011, 11:24 a.m. | |
| 9 | -3.46% | leonidas/corner_catches256/2 | Medium/2/leonidas/corner_catches256/2 | 42.92% / 44.32% | Oct. 25, 2011, 11:24 a.m. | |
| 10 | -3.43% | leonidas/dt_hogs10/1 | Same heuristic different parameters | Medium/2/leonidas/dt_hogs10 | 42.97% / 44.35% | Oct. 25, 2011, 11:24 a.m. |
| 11 | -3.30% | cdubout/hog/6 | Histogram of oriented gradients (orientation discretized in 12 bins) taken at random positions and scales. Strongly inspired from francoisfleuret/zk_v2. | Medium/2/cdubout/hog/6 | 42.91% / 44.48% | Oct. 25, 2011, 11:23 a.m. |
| 12 | -3.26% | leonidas/lopd/13 | Pooled gradients taken at positions on a circular grid (Daisy-like). | Medium/2/leonidas/lopd/13 | 43.25% / 44.52% | Oct. 25, 2011, 11:24 a.m. |
| 13 | -2.92% | leonidas/catches/4 | Another codebook based descriptor | Medium/2/leonidas/catches/4 | 43.84% / 44.86% | Oct. 25, 2011, 11:24 a.m. |
| 14 | -2.77% | cdubout/hogblurred/4 | Strongly blurred gradient magnitude images for 8 different gradient orientations (also rescales the image to 25% to save on features). | Medium/2/cdubout/hogblurred/4 | 43.24% / 45.01% | Oct. 25, 2011, 11:23 a.m. |
| 15 | -2.66% | francoisfleuret/zk/4 | Computes proportion of edges over rectangles picked at random. | Medium/2/francoisfleuret/zk/4 | 43.70% / 45.12% | Oct. 25, 2011, 11:23 a.m. |
| 16 | -2.57% | cdubout/patches/3 | Histogram of the (maximum and minimum) correlation with a dictionary of patches. | Medium/2/cdubout/patches/3 | 43.59% / 45.21% | Oct. 25, 2011, 11:23 a.m. |
| 17 | -2.56% | leonidas/scratches/4 | Yet another codebook descriptor. | Medium/2/leonidas/scratches/4 | 43.79% / 45.22% | Oct. 25, 2011, 11:24 a.m. |
| 18 | -2.05% | kanma/rgbhistogram/1 | This heuristic computes the histogram of the region of interest, in RGB format. The RGB color-space is segmented in 512 parts (8 per channel). | Medium/2/kanma/rgbhistogram | 44.49% / 45.73% | Oct. 25, 2011, 11:24 a.m. |
| 19 | -1.99% | gillesblanchard/blockwise2/1 | Medium/2/gillesblanchard/blockwise2 | 44.33% / 45.79% | Oct. 25, 2011, 11:23 a.m. | |
| 20 | -1.95% | francoisfleuret/chamferzk/4 | For each pixel, computes 8 Chamfer distances to as many edge orientations. | Medium/2/francoisfleuret/chamferzk/4 | 44.28% / 45.83% | Oct. 25, 2011, 11:23 a.m. |
| 21 | -1.71% | cdubout/viola_jones/4 | The rectangle features used by Viola and Jones in their famous IJCV paper "Robust Real-time Object Detection". | Medium/2/cdubout/viola_jones/4 | 44.35% / 46.07% | Oct. 25, 2011, 11:23 a.m. |
| 22 | -1.66% | cdubout/hough/3 | Computes the linear Hough transform of the region of interest to detect lines. | Medium/2/cdubout/hough/3 | 44.65% / 46.12% | Oct. 25, 2011, 11:23 a.m. |
| 23 | -1.61% | cdubout/fourier/4 | Fourier transform heuristic. Computes the 2D FFT of the region of interest to convert it to the frequency domain. | Medium/2/cdubout/fourier/4 | 44.76% / 46.17% | Oct. 25, 2011, 11:22 a.m. |
| 24 | -1.60% | kanma/blobs/1 | This heuristic produces a 'blobbed' version of the (RGB) region of interest. | Medium/2/kanma/blobs | 44.70% / 46.18% | Oct. 25, 2011, 11:23 a.m. |
| 25 | -1.41% | leonidas/laws/2 | A texture descriptor using 3x3 Laws masks | Medium/2/leonidas/laws/2 | 44.87% / 46.37% | Oct. 25, 2011, 11:24 a.m. |
| 26 | -1.40% | gillesblanchard/blockwise/1 | Differences of blockwise averages at different scales on the ROI (Haar wavelets à la Viola and Jones) | Medium/2/gillesblanchard/blockwise | 45.07% / 46.38% | Oct. 25, 2011, 11:23 a.m. |
| 27 | -1.30% | jena/radialedgesstats/4 | Computes several edge-based statistics along the radial directions. | Medium/2/jena/radialedgesstats/4 | 45.02% / 46.48% | Oct. 25, 2011, 11:23 a.m. |
| 28 | -1.28% | francoisfleuret/boxedaverages/3 | Computes the average gray level over rectangular windows picked at random. | Medium/2/francoisfleuret/boxedaverages/3 | 45.10% / 46.50% | Oct. 25, 2011, 11:23 a.m. |
| 29 | -1.16% | cdubout/haar/1 | Haar transform heuristic. Computes the 2D haar transform of the region of interest (all levels). | Medium/2/cdubout/haar | 44.91% / 46.62% | Oct. 25, 2011, 11:22 a.m. |
| 30 | -1.16% | dubecmar/normalization/1 | Heuristic to stretch the range of intensities include intensity normalization | Medium/2/dubecmar/normalization | 45.27% / 46.62% | Oct. 25, 2011, 11:23 a.m. |
| 31 | -1.12% | cdubout/identity/1 | Same as the mash/identity heuristic except that it returns RGB values instead of grayscale. | Medium/2/cdubout/identity | 44.78% / 46.66% | Oct. 25, 2011, 11:23 a.m. |
| 32 | -1.05% | mash/identity/1 | This heuristic is part of the MASH SDK | Medium/2/mash/identity | 44.98% / 46.73% | Oct. 25, 2011, 11:25 a.m. |
| 33 | -0.58% | cdubout/segmentation8/1 | Segment the gray-scale image intensities into 8 bins with equal number of pixels in each. Returns the index of the bins in which a new pixel falls. | Medium/2/cdubout/segmentation8 | 45.32% / 47.20% | Oct. 25, 2011, 11:23 a.m. |
| 34 | -0.58% | andrebeinrucker/simple_pixel_difference/1 | Checks whether the difference of a pixel value to the value of the neighbor pixel is higher than a threshold. | Medium/2/andrebeinrucker/simple_pixel_difference | 45.58% / 47.20% | Oct. 25, 2011, 11:22 a.m. |
| 35 | -0.54% | kanma/grayscalehistogram/1 | This heuristic computes the histogram of the region of interest, in grayscale format | Medium/2/kanma/grayscalehistogram | 45.81% / 47.24% | Oct. 25, 2011, 11:24 a.m. |
| 36 | -0.50% | leonidas/simple_color_hist/1 | Average RGB values | Medium/2/leonidas/simple_color_hist | 45.88% / 47.28% | Oct. 25, 2011, 11:24 a.m. |
| 37 | -0.44% | mash/meanthreshold/1 | This heuristic is part of the MASH SDK | Medium/2/mash/meanthreshold | 45.89% / 47.34% | Oct. 25, 2011, 11:25 a.m. |
| 38 | -0.40% | kanma/fartherpixel/1 | This heuristic looks at 9x9 blocks of grayscale pixels and returns the index of the one which is the farther (intensity-wise) of the center one. | Medium/2/kanma/fartherpixel | 45.71% / 47.38% | Oct. 25, 2011, 11:24 a.m. |
| 39 | -0.36% | jena/randomprojections/1 | Computes a set of random projections of the gray-scale ROI. | Medium/2/jena/randomprojections | 46.31% / 47.42% | Oct. 25, 2011, 11:23 a.m. |
| 40 | -0.33% | lmerchante/symmetryax/3 | Calculates a score of the ROI axial symmetry from 0º to 180º degrees rotations of the axe of symmetry | Medium/2/lmerchante/symmetryax/3 | 45.86% / 47.45% | Oct. 25, 2011, 11:25 a.m. |
| 41 | -0.31% | cdubout/segmentation2/1 | Segment the gray-scale image pixels into 2 categories, the ones above the average intensity, and the ones below. | Medium/2/cdubout/segmentation2 | 45.95% / 47.47% | Oct. 25, 2011, 11:23 a.m. |
| 42 | -0.28% | francoisfleuret/average/1 | Return the average level of the pixel gray levels. | Medium/2/francoisfleuret/average | 46.33% / 47.50% | Oct. 25, 2011, 11:23 a.m. |
| 43 | -0.25% | kanma/closerpixel/1 | This heuristic looks at 9x9 blocks of grayscale pixels and returns the index of the one which is the closer (intensity-wise) of the center one. | Medium/2/kanma/closerpixel | 45.69% / 47.53% | Oct. 25, 2011, 11:24 a.m. |
| 44 | -0.07% | lmerchante/edgessymmetryax/1 | Evaluates the axial symmetry of the edges detected in the image | Medium/2/lmerchante/edgessymmetryax | 46.22% / 47.71% | Oct. 25, 2011, 11:24 a.m. |
| 45 | 0.01% | lmerchante/thresholdgradient/1 | Calculates the gradient of a thresholded image. It's a kind of very simple edge detector. | Medium/2/lmerchante/thresholdgradient | 46.29% / 47.79% | Oct. 25, 2011, 11:25 a.m. |
| 46 | 0.11% | lmerchante/gradient/1 | 2D gradient calculation as the addition of 8 different gradients regarding 8 orientations at every pixel | Medium/2/lmerchante/gradient | 46.21% / 47.89% | Oct. 25, 2011, 11:24 a.m. |
| 47 | 0.35% | lmerchante/blurconv/2 | This heuristic evaluates the convolutiuon of the ROI with a 5x5 mean MASK. | Medium/2/lmerchante/blurconv/2 | 45.81% / 48.13% | Oct. 25, 2011, 11:24 a.m. |
| 48 | 0.45% | lmerchante/symmetript/1 | Calculates an score of symmetry for every pixel in the image | Medium/2/lmerchante/symmetript/1 | 46.29% / 48.23% | Oct. 25, 2011, 11:24 a.m. |