THINKING THROUGH DRAWING: PRACTICE INTO KNOWLEDGE

THINKING THROUGH DRAWING: PRACTICE INTO KNOWLEDGE 77

Ruben Coen-Cagli

movements and the experimental scanpaths. Figure
4 summarizes scanpath similarity values obtained
by comparing either the drawing task (Fig. 11(a))
or free viewing scanpaths (Fig. 11(b)), against: 1)
(red bars) the sequence of gaze-points generated by
our model, which implements the edge-following;
2) (blue bars) scanpaths generated by the saliency-
based algorithm; and 3) (green bars) random
scanpaths (averaged over 10000 samples). The full
model performs significantly better than chance, as
well as better than a purely saliency-based model,
in the drawing task; conversely, the control experi-
ment showed that both versions of the model were,
on average, as poor as chance in capturing free
viewing scanpaths. See Figure 4.

Discussion
We studied eye movements during copy draw-

ing and observed a local processing bias and “edge-

following” scanpaths, markedly different from the

patterns observed in the free-viewing condition.

With the aid of a computational model that com-

putes low-level image features, and also learns the

coordination of eye- and hand-related variables, we

showed that the latter piece of information was nec-

essary to account for the observed data in the copy-

drawing experiment; conversely, neither image

features nor visuomotor coordination could explain

the free-viewing scanpaths better than chance. Our

observations suggest that the direct transformation

of sensory inputs (visual and proprioceptive) into

motor plans (for the eye and the hand) is a core pro-

cess of the activity of copy-drawing in non-artists.

Future work should address how the results

presented here differ between naive and subjects

and experienced artists. This could include fac-

Figure 4. Left: example scanpaths obtained in the experiments, and by the full model and the salience model.
Right-top: The similarity, as measured by the Levenshtein distance, between experimental scanpaths in the
drawing task and those simulated by our model (blue), by a saliency-based algorithm (green), and randomly
generated (red); bars denote the values for each subject, while triangles denote mean value and standard
deviation across subjects. Right-bottom: same as above, but with human data obtained in the free viewing
condition.