Exploit Bounding Box Annotations for Multi-label Object Recognition

Convolutional neural networks (CNNs) have shown great performance as general feature representations for object recognition applications. However, for multi-label images that contain multiple objects from different categories, scales and locations, global CNN features are not optimal. In this paper, we incorporate local information to enhance the feature discriminative power. In particular, we first extract object proposals from each image. With each image treated as a bag and object proposals extracted from it treated as instances, we transform the multi-label recognition problem into a multi-class multi-instance learning problem. Then, in addition to extracting the typical CNN feature representation from each proposal, we propose to make use of ground-truth bounding box annotations (strong labels) to add another level of local information by using nearest-neighbor relationships of local regions to form a multi-view pipeline. The proposed multi-view multi-instance framework utilizes both weak and strong labels effectively, and more importantly it has the generalization ability to even boost the performance of unseen categories by partial strong labels from other categories. Our framework is extensively compared with state-of-the-art hand-crafted feature based methods and CNN based methods on two multi-label benchmark datasets. The experimental results validate the discriminative power and the generalization ability of the proposed framework. With strong labels, our framework is able to achieve state-of-the-art results in both datasets.

HST Rotational Spectral Mapping of Two L-Type Brown Dwarfs: Variability In and Out of Water Bands Indicates High-Altitude Haze Layers

We present time-resolved near-infrared spectroscopy of two L5 dwarfs, 2MASS J18212815+1414010 and 2MASS J15074759-1627386, observed with the Wide Field Camera 3 instrument on the Hubble Space Telescope (HST). We study the wavelength dependence of rotation-modulated flux variations between 1.1 $mu$m and 1.7 $mu$m. We find that the water absorption bands of the two L5 dwarfs at 1.15 $mu$m and 1.4 $mu$m vary at similar amplitudes as the adjacent continuum. This differs from the results of previous HST observations of L/T transition dwarfs, in which the water absorption at 1.4 $mu$m displays variations of about half of the amplitude at other wavelengths. We find that the relative amplitude of flux variability out of the water band with respect to that in the water band shows a increasing trend from the L5 dwarfs toward the early T dwarfs. We utilize the models of Saumon & Marley (2008) and find that the observed variability of the L5 dwarfs can be explained by the presence of spatially varying high-altitude haze layers above the condensate clouds. Therefore, our observations show that the heterogeneity of haze layers - the driver of the variability - must be located at very low pressures, where even the water opacity is negligible. In the near future, the rotational spectral mapping technique could be utilized for other atomic and molecular species to probe different pressure levels in the atmospheres of brown dwarfs and exoplanets and uncover both horizontal and vertical cloud structures.