In visual scenes, the presence of corner elements greatly simplifies object recognition [1]. It has been suggested that cortical area V4 is responsible for the processing of these elements, since neurons exhibit selectivity for different curvatures [2]. However, V4 response properties and their contribution to contour integration or shape completion are still far from being understood. We investigated visually-evoked responses of macaque V4 neurons to sparse random stimuli, composed of angled segments pointing in two directions like clock hands (Fig 1a). Recordings were obtained using a 96-channel Utah array [3]. Reverse correlation on multi-unit activity MUA, entire spiking activity ESA, and broad/narrowband (nb, 35–51Hz) local field potentials LFP was used to compute spatio-temporal response kernels. We applied a careful statistical analysis to reduce noise and determine contiguous receptive field (RF) profiles significantly deviating from zero. Finally, we tested the kernels’ predictive power in a linear-nonlinear model by computing the Pearson correlation between model and recorded responses. We focused first on kernels estimated from stimulus-onset responses, averaged over different angle configurations. RFs were well localized in time (peak at 50–100ms, followed by response of opposite polarity for ESA/LFP) and in space, with consistent size increase across signal features (smallest in MUA, +70% in ESA/nbLFP, +140% in LFP). Most spatial RFs were unimodal, but LFPs exhibited on/off subfields. Predicting the neural response on test data was best for LFPs, reaching correlations of 0.27. Response selectivity to different angle configurations was quantified by assessing significant deviation of the neurons’ spatial RFs for specific configurations from their average. Selectivity was highest in nbLFPs around 80ms for 41 out of 88 channels. Interestingly, at 110ms 27 channels were significantly non-selective, i.e. configuration-specific RFs deviated much less from their average than expected from the noise level. However, it was not possible to predict the neural response from the estimated kernels. This might partly be due to overfitting since kernels retained size, but training data reduced to 1/28 for shape-selective estimation. In summary, V4 response selectivity to corner elements is less pronounced than expected, with RFs exhibiting a complex structure (Fig 1b-e) indicating a much stronger influence from non-linear effects like surround interactions.