It would certainly be possible to envision a shape processing system that balked at outlines and used only surface edges. As introductory drawing teachers often stress, outlines are ecologically anomalous.
CROCKS NEWSLETTER - 30 SEPTEMBER
Within biological vision, some perceptual processes seem to treat outlines and surface edges differently, as in perceptual completion [ 33 ]. Humans can see shape given by ordinary surface edges or outlines, but if deep networks cannot utilize the latter, it does not necessarily imply that surfaces edges do not play some role in classification in conjunction with texture features. Humans' fluent use of outlines to see object shape indicates the strong role of shape representation in human object recognition, and human interpretation of forms in outlines probably connects naturally to some stage of perceptual representation [ 34 ].
It appears that DCNNs differ from human processors in that they have little or no linkage between shape properties embodied in outlines and the classification labels in the output layer. These issues may relate to an important factor not yet mentioned. Figure-ground assignment, or equivalently, assignment of border ownership at occluding edges, is a well-known feature of human perceptual organization [ 35 , 36 ].
It appears that outlines, especially closed outlines, are interpreted in human vision as owning their borders. The enclosed area is taken to be the bounded object.
DCNNs do not have an obvious way of representing figure vs. These seem to be more explicitly representational aspects of human perceptual processing.
At least some of the problem with outlines may involve figure-ground issues. On the other hand, results for a few of the images presented in Exp. If the networks were only treating the black outlines as the figure, these objects should have below-chance probability, as they do not have thin, black forms. Of course, we do not mean to imply that a DCNN employs any consistent approach or strategy; any good predictor from any of the many filters in the network, may influence the outcomes toward a correct classification.
Perhaps these images have certain local features that can be extracted and further facilitate the recognition. DCNNs may not capture global shape, but may pick up some relatively local shape features, a possibility we discuss further in connection with later results. Experiments 2 and 3 found little evidence that deep convolutional networks access global shape in object recognition tasks. These results may seem surprising, as some recent reports have suggested that DCNNs do possess some shape classification abilities.
Kubilius et al. On the one hand, this marks a divergence from human performance, which is largely unaffected by the removal of color and inner surface gradient information from most objects. In contrast, our findings about recognition for glass objects and line drawings provided almost no evidence that shape representations are used for classification in DCNNs. In Experiment 4, we tried to replicate Kubilius et al. The same 40 object silhouettes found on the internet and used in Experiment 1 were used in Experiment 4, this time without any texture substitution.
All images consisted of a single black figure on a white background. Half of the images were artifacts, and half were animals. The black figures were silhouettes of object drawings, rather than being taken from photographs of real objects with their textures removed, so some contour information that would typically be in a natural instance of the object was abstracted in the silhouette images. See Fig 18 for examples. We also tested the network on the same 40 silhouettes with white figures on a black background and red figures on a white background to measure the influence of surface color on network classification performance.
For the 40 black silhouettes on a white background, VGG and AlexNet correctly classified 20 and 15 of the 40 presented images in their top-five classifications , respectively. Figs 19 — 22 shows the results for VGG Performance was worse for images with white figures on black grounds, where the network classified seven of the 40 images correctly, and for images with red figures on white grounds, where the network classified nine of the 40 images correctly.
The leftmost column shows the image presented to VGG The second column from the left shows the correct object label and the classification probability produced for that label. The results from Experiment 4 are largely consistent with the findings reported by Kubilius et al. Performance was notably worse for white-on-black and red-on-white figures than for black-on-white figures. One reason for this might be that there are more canonically black objects in the training set than white or red.
For example, the cannon is correctly classified when presented as a black figure, but incorrectly classified when presented as a red figure. Another reason networks might be better at classifying black figures is that they more closely resemble photographic images that were used in network training. Objects will appear very dark or even black if they are between the camera and a bright light, as in actual silhouettes. Possibly, exposure to training examples like these makes the network more likely to accept dark figures as instances of an object, even one that is not canonically dark.
The differences in network performance across these three testing sets points to the strong influence of surface information in classification. For humans, a homogenous surface texture would likely not be considered at all in recognition, as the visual system would recognize that there is not enough surface information present to be diagnostic. The network makes no such evaluation and remains highly sensitive to such cues.
Chamber Music in Wales
Regarding shape, this experiment showed a clear contribution of contour properties in classification of object silhouettes. Within a given display set, all of the test displays shared the same coloration; therefore, all differences in classification responses from the DCNNs involved contour information. Performance at this level demands explanations that go beyond a simple conclusion that DCNNs do or do not process object shape. If DCNNs had access to global shape information, we might have expected their performance to be similar to humans, readily producing accurate classifications for all displays.
The results also contain some rather conspicuous failures to process overall shape. For the porcupine display, for example, VGG gave, as its top choice, "bald eagle". For "lion", AlexNet's top choice was "goose", with high confidence, more than twice the probability of any other response.
- From Monastery to Hospital: Christian Monasticism and the Transformation of Health Care in Late Antiquity!
- Intervention: How America Became Involved in Vietnam!
- Navigation menu;
- Stripes and related phenomena!
- Selling Women: Prostitution, Markets, and the Household in Early Modern Japan.
- Plants for environmental studies;
These results contain important information regarding processing of overall shape. For humans, at least, the lion display is both recognizable as having the shape of a lion and is clearly not shaped at all like a goose. It could be pointed out that failure to give a certain label may merely indicate that the particular shape captured in silhouette may have had a particular vantage point that was uncharacteristic of examples in the training set. The implications for global shape processing here, however, hinge less on the selection of the correct name than on the incorrect answers furnished.
These results suggest that overall shape is elusive in DCNN responses, but also that something relating to shape allows success in some of the cases tested. Perhaps a deeper analysis of what is meant by shape is needed to understand both the successes and failures of DCNNs. We consider this in successive steps below.dextviburno.cf
First, why is classification so much better for object silhouettes than for glass figures and shape outlines? We have already commented that the ability to use outlines as depicting shape, although significant in human perception, is not a necessary condition for a DCNN to be shape processor. To confirm that the difference between network performance on silhouettes and outlines was not item specific, we sampled the outline of the 40 silhouettes and tested the network on outline images of the stimuli in Experiment 4.
The results closely matched those reported in Experiment 3—the network only classified three of the 40 images in its top-five selections. What about the better performance of black silhouettes over glass objects? One reason may be that silhouettes successfully reduce distracting texture information, i. Glass objects may have contained more misleading surface information than black silhouettes. There are more objects with glass or other transparent or reflective surfaces in the categories on which the networks were trained than there are uniformly black objects.
As mentioned, silhouettes also have the advantage of potentially looking similar to some photographs of the objects during network training if the photographs were taken at sunset or with a bright light behind the object. Another important factor contributing to the differences in the previous three experiments is the involvement of figure-ground segmentation.
Silhouettes are likely easier to classify than glass objects because the surface information provided contains no internal contours. Although figure-ground segmentation is an important part of human shape processing, it is likely that DCNNs produce object classifications with natural images without performing any explicit figure-ground segmentation. In human perception, bounding contours are defining of shape, and shape descriptions are conferred based on the bounding contours of segmented objects [ 38 ].
Without any figure-ground segmentation mechanism, all contours in training examples probably have equal status. Nothing designates a bounding contour relevant to overall shape, as opposed to contour information that may be part of surface texture, or noise, etc. With displays stripped of all contour information except for bounding contours, the networks do better. This suggests that the networks must have used some information about the forms of objects to achieve the performance observed in Experiment 4, even though performance with silhouettes still falls far short of classification that humans readily do with shape.
By comparison with the earlier experiments, it also suggests that important bounding contour information is more influential when no other contours are present.
- Silhouette No. 9 in B Major from "Silhouettes", Op. 8 | Sheet Music Now.
- silhouette no 2 in d flat major from silhouettes op 8 Manual.
- Higher Engineering Mathematics (5th Edition)!
- Videos matching Antonín Dvořák Silhouettes Op.8, Radoslav Kvapil | Revolvy.
What use is made of contour information? An important insight may be provided by the results for the black bear silhouette. The silhouette is substantially simplified such that key points of concavity along a bear outline are connected, mostly by straight lines.
- Send us a message.
- List of compositions by Max Reger!
- Antonin Dvorak. MIDI (free download) & MIDI/ZIP.
- By Composer.
- Representation as Violence: The Art of Kara Walker?
- Technologies of Gender: Essays on Theory, Film, and Fiction!
- A Methodology for Determining Air Force Deployment Requirements (Supporting Air and Space Expeditionary Forces)!
Attneave observed that humans can robustly recognize objects whose contour has been changed significantly at a local level, provided that at the global level the spatial relationships between important points along the contour are preserved. Deep networks do not appear to have the same capabilities. We suspect that they are doing essentially the reverse of humans with regard to global and local aspects of shape.
We refer to this as the local contour feature hypothesis.