| | <Poster Width="1734" Height="1041"> |
| | <Panel left="9" right="115" width="526" height="277"> |
| | <Text>Introduction & Motivation</Text> |
| | <Text>unsolved problems, and bridge computer and human</Text> |
| | <Text>vision, we define a battery of 5 tests that measure the</Text> |
| | <Text>gap between human and machine performances in</Text> |
| | <Text>several dimensions.- Here, to help focus research efforts onto the hardest</Text> |
| | <Text>design new experiments to understand mechanisms</Text> |
| | <Text>of human vision, and to reason about its failure.</Text> |
| | <Text>Cases where humans are better inspire computational</Text> |
| | <Text>researchers to learn from humans.- Cases where machines are superior motivate us to</Text> |
| | <Text>or personal robots), perfect accuracy is not necessarily</Text> |
| | <Text>the goal; rather, having the same type of behavior (e.g.,</Text> |
| | <Text>failing in cases where humans fail too) is favorable.In some applications (e.g., human-machine interaction</Text> |
| | <Figure left="329" right="148" width="203" height="240" no="1" OriWidth="0.373988" OriHeight="0.322163 |
| | " /> |
| | </Panel> |
| |
|
| | <Panel left="11" right="397" width="524" height="393"> |
| | <Text>Test 1: Scene Recognition</Text> |
| | <Figure left="19" right="423" width="165" height="147" no="2" OriWidth="0.254335" OriHeight="0.169794 |
| | " /> |
| | <Figure left="183" right="419" width="171" height="153" no="3" OriWidth="0.245665" OriHeight="0.167113 |
| | " /> |
| | <Figure left="356" right="422" width="174" height="149" no="4" OriWidth="0.253179" OriHeight="0.169348 |
| | " /> |
| | <Text>A & B) Scene classification accuracy over 6-, 8- and 15-CAT datasets. Error bars represent the standard error of the mean over 10 runs. Naive</Text> |
| | <Text>chance is simply set by the size of the largest class. All models work well above chance level. C) Top: animal vs. non-animal (distractor images)</Text> |
| | <Text>classification. Bottom: classification of target images. 4-way classification is only over target scenes (and not distractors).</Text> |
| | <Text>- A & B: We find that HOG, SSIM, texton, denseSIFT, LBP,</Text> |
| | <Text>and LBHPF outperform other models (accuracy above 70%).</Text> |
| | <Text>We note that spatial feature integration (i.e., x_pyr for the</Text> |
| | <Text>model x) enhances accuracies.</Text> |
| | <Text>- C: Animal vs. Non-Animal: All models perform above 70%,</Text> |
| | <Text>except tiny image. Human accuracy here is about 80%. Inter-</Text> |
| | <Text>estingly, some models exceed human performance here.</Text> |
| | <Text>SUN dataset: Models that performed well on small datasets</Text> |
| | <Text>(although they degrade heavily) still rank on top. GIST model</Text> |
| | <Text>works well here (16.3%) but below top contenders: HOG, tex-</Text> |
| | <Text>ton, SSIM, denseSIFT, and LBP (or their variants). Models</Text> |
| | <Text>ranking at the bottom, in order, are tiny image, line hist, geo</Text> |
| | <Text>color, HMAX, and geo map8x8.</Text> |
| | <Figure left="302" right="611" width="235" height="147" no="5" OriWidth="0.37341" OriHeight="0.180965 |
| | " /> |
| | <Text>Performances and correlations on SUN dataset. We randomly</Text> |
| | <Text>chose n = { 1, 5, 10, 20, 50} images per class for training and 50 for test.</Text> |
| | </Panel> |
| |
|
| | <Panel left="11" right="795" width="522" height="234"> |
| | <Text>Learned Lessons</Text> |
| | <Text>1) Models outperform humans in rapid categorization tasks, indicating that discriminative informa-</Text> |
| | <Text>tion is in place but humans do not have enough time to extract it. Models outperform humans on</Text> |
| | <Text>jumbled images and score relatively high in absence of (less) global information.</Text> |
| | <Text>2) We find that some models and edge detection methods are more efficient on line drawings and</Text> |
| | <Text>edge maps. Our analysis helps objectively assess the power of edge detection algorithms to ex-</Text> |
| | <Text>tract meaningful structural features for classification, which hints toward new directions.</Text> |
| | <Text>3) While models are far from human performance over object and scene recognition on natural</Text> |
| | <Text>scenes, even classic models show high performance and correlation with humans on sketches.</Text> |
| | <Text>4) Consistent with the literature, we find that some models (e.g., HOG, SSIM, geo/texton, and</Text> |
| | <Text>GIST) perform well. We find that they also resemble Fighumans better.</Text> |
| | <Text>5) Invariance analysis shows that only sparseSIFT and geo_color are invariant to in-plane rotation</Text> |
| | <Text>with the former having higher accuracy (our 3rd test). GIST, a model of scene recognition works</Text> |
| | <Text>better than many models over both Caltech-256 and Sketch datasets.</Text> |
| | </Panel> |
| |
|
| | <Panel left="540" right="116" width="490" height="915"> |
| | <Text>Test 2: Recognition of Line Drawings and Edge Maps</Text> |
| | <Text>Line Drawings</Text> |
| | <Text>- Scenes were were presented</Text> |
| | <Text>to subjects for 17-87 ms</Text> |
| | <Text>in a 6-alternative force choice</Text> |
| | <Text>task (human acc= 77.3%).</Text> |
| | <Text>- On color images, geo_color,</Text> |
| | <Text>sparseSIFT, GIST, and SSIM</Text> |
| | <Text>showed the highest correla-</Text> |
| | <Text>tion (all with classification ac-</Text> |
| | <Text>curacy ≥ 75%), while tiny</Text> |
| | <Text>images, texton, LBHF, and</Text> |
| | <Text>LBP showed the least. Over</Text> |
| | <Text>the SUN dataset, HOG,</Text> |
| | <Text>denseSIFT, and texton</Text> |
| | <Text>showed high correlation with</Text> |
| | <Text>human CM.</Text> |
| | <Text>- It seems that those models</Text> |
| | <Text>that take advantage of re-</Text> |
| | <Text>gional histogram of features</Text> |
| | <Text>(e.g., denseSIFT, GIST, geo_</Text> |
| | <Text>x; x=map or color) or heavily</Text> |
| | <Text>rely on edge histograms</Text> |
| | <Text>(texton and HOG) show</Text> |
| | <Text>higher correlation with</Text> |
| | <Text>humans on color images</Text> |
| | <Text>(although low in magnitude).</Text> |
| | <Text>- Over line drawings: As </Text> |
| | <Text>images, geo_color, SSIM,</Text> |
| | <Text>and sparseSIFT correlate </Text> |
| | <Text>with humans.To our surprise,</Text> |
| | <Text>geo_color worked well.</Text> |
| | <Figure left="714" right="168" width="318" height="233" no="6" OriWidth="0.383815" OriHeight="0.219839 |
| | " /> |
| | <Text> Human-model agreement on the 6-CAT dataset. See our paper</Text> |
| | <Text>and its supplement for confusion matrices of models.</Text> |
| | <Figure left="711" right="437" width="321" height="101" no="7" OriWidth="0.383815" OriHeight="0.219839 |
| | " /> |
| | <Text> Geometric map: ground, pourous, sky, and vertical regions.</Text> |
| | <Figure left="712" right="570" width="319" height="56" no="8" OriWidth="0.374566" OriHeight="0.0469169 |
| | " /> |
| | <Text> Edge maps for a sample image .</Text> |
| | <Text>Edge Maps</Text> |
| | <Figure left="548" right="670" width="479" height="86" no="9" OriWidth="0.772832" OriHeight="0.108579 |
| | " /> |
| | <Text> Scene classification results using edge detected images over 6-CAT dataset. Canny edge detector</Text> |
| | <Text>leads to best accuracies followed by the log and gPb methods.</Text> |
| | <Text>- A majority of models perform > 70% on line</Text> |
| | <Text>drawings which is higher than human perfor-</Text> |
| | <Text>mance (similar pattern on images with</Text> |
| | <Text>human=77.3% and models > 80%).</Text> |
| | <Text>- SVM trained on images and tested on line</Text> |
| | <Text>drawings: Some models (e.g., line hists, GIST,</Text> |
| | <Text>geo map, sparseSIFT) better generalize to </Text> |
| | <Text>drawings.</Text> |
| | <Text>SVM trained on line drawings and tested on</Text> |
| | <Text>edge maps: Surprisingly, averaged over all</Text> |
| | <Text>models, Sobel and Canny perform better than</Text> |
| | <Text>gPb. GIST, line hists, and HMAX were the most</Text> |
| | <Text>successful models using all edge detection</Text> |
| | <Text>methods. sparseSIFT, LBP, geo_color, and</Text> |
| | <Text>geo_texton were the most affected ones.</Text> |
| | <Text>- Models using Canny technique achieved the</Text> |
| | <Text>best scene classification accuracy.</Text> |
| | <Figure left="773" right="781" width="258" height="211" no="10" OriWidth="0.373988" OriHeight="0.241734 |
| | " /> |
| | <Text> Top: training a SVM from color photographs and testing on</Text> |
| | <Text>line drawings, gPb edge maps, and inverted (FL) images. Bottom: SVM</Text> |
| | <Text>trained on line drawings and applied to edge maps.</Text> |
| | </Panel> |
| |
|
| | <Panel left="1035" right="115" width="349" height="396"> |
| | <Text>Test 3: Invariance Analysis</Text> |
| | <Figure left="1039" right="142" width="328" height="238" no="11" OriWidth="0.371676" OriHeight="0.230563 |
| | " /> |
| | <Text> d values over original, 90 o , and 180 o rotated animal images.</Text> |
| | <Text>- A majority of models are invariant to scaling while few are drasti-</Text> |
| | <Text>cally affected with a large amount of scaling (e.g., siagianItti07,</Text> |
| | <Text>SSIM, line hists, and sparseSIFT).</Text> |
| | <Text>- Interestingly, LBP here shows a similar pattern as humans across</Text> |
| | <Text>four stimulus categories (i. e., max for head, min for close body).</Text> |
| | <Text>- Some models show higher similarity to human disruption over the</Text> |
| | <Text>four categories of the animal dataset: sparseSIFT, SSIM, and HOG.</Text> |
| | </Panel> |
| |
|
| | <Panel left="1387" right="117" width="333" height="394"> |
| | <Text>Test 4: Local vs. Global Information</Text> |
| | <Figure left="1397" right="141" width="326" height="212" no="12" OriWidth="0.372254" OriHeight="0.188561 |
| | " /> |
| | <Text> Correlation and classification accuracy over jumbled images.</Text> |
| | <Text>As expected, models based on histograms are less influ-</Text> |
| | <Text>enced (e.g ., geo color, line hist, HOG, texton, and LBP).</Text> |
| | <Text>- Models correlate higher with humans over scenes (OSR and</Text> |
| | <Text>ISR) than objects, and better on outdoor scenes than indoors.</Text> |
| | <Text>- Some models, which use global feature statistics, show high</Text> |
| | <Text>correlation only on scenes but very low on objects (e.g.,</Text> |
| | <Text>GIST, texton, geo map, and LBP), since they do not capture</Text> |
| | <Text>object shape or structure.</Text> |
| | </Panel> |
| |
|
| | <Panel left="1032" right="516" width="689" height="361"> |
| | <Text>Test 5: Object Recognition</Text> |
| | <Text>- On Caltech-256,</Text> |
| | <Text>HOG achieves the</Text> |
| | <Text>highest accuracy</Text> |
| | <Text>about 33.28% fol-</Text> |
| | <Text>lowed by SSIM,</Text> |
| | <Text>texton, and dense</Text> |
| | <Text>SIFT.</Text> |
| | <Text>GIST which is spe-</Text> |
| | <Text>cifically designed for</Text> |
| | <Text>scene categorization</Text> |
| | <Text>achieves 27.4% accu-</Text> |
| | <Text>racy, better than some</Text> |
| | <Text>models specialized</Text> |
| | <Text>for object recognition</Text> |
| | <Text>(e.g., HMAX).</Text> |
| | <Figure left="1170" right="544" width="276" height="217" no="13" OriWidth="0.388439" OriHeight="0.258713 |
| | " /> |
| | <Figure left="1444" right="541" width="276" height="220" no="14" OriWidth="0.387861" OriHeight="0.258266 |
| | " /> |
| | <Text> Left: Object recognition performance on Caltech-256 dataset. Right: Recognition rate and correlations on Sketch dataset.</Text> |
| | <Text>On sketch images, the shogSmooth model, specially designed for recognizing sketch images, outperforms others</Text> |
| | <Text>(acc=57.2%). Texton histogram and SSIM ranked second and fourth, respectively. HMAX did very well (in contrast to</Text> |
| | <Text>Caltech-256), perhaps due to its success in capturing edges, corners, etc.</Text> |
| | <Text>- Overall, models did much better on sketches than on natural objects (results are almost 2 times higher than the Caltech-</Text> |
| | <Text>256). Here, similar to the Caltech-256, features relying on geometry (e.g., geo_map) did not perform well.</Text> |
| | </Panel> |
| |
|
| | <Panel left="1034" right="883" width="683" height="149"> |
| | <Text>Summary</Text> |
| | <Figure left="1081" right="884" width="621" height="128" no="15" OriWidth="0.719653" OriHeight="0.119303 |
| | " /> |
| | <Text> Classification results corresponding to 50 training and (50 over SUN and remaining images over Caltech-256 and Sketch) testing images per class</Text> |
| | <Text>Animal vs. non-Animal corresponds to classification of 600 target vs. 600 distractor images . Top three models on each dataset are highlighted in red.</Text> |
| | </Panel> |
| |
|
| | </Poster> |
| |
|
