How do ants use encode & identify natural panoramic scenes?

全景场景对于许多导航动物来说很重要，因为它们提供了方向信号(家在山的左边)，指南针方向(山谷从北到南)，以及粗略的地方感(我们快到家了)。但大片的自然场景是如何编码或识别的，几乎是未知的。我们将在蚂蚁身上探索这一主题，因为它是研究昆虫模式视觉的一种新方法，更广泛地说，因为低分辨率和相对简单的视觉系统处理自然场景的方式涉及当前神经科学、计算机视觉和机器人学的领域。行走的蚂蚁是这类研究的好昆虫，因为它们在识别场景和接近模式时的扫描和查看运动可以以高分辨率进行监测，从而给出蚂蚁体轴的方向和位置，使我们能够推断场景是如何在视网膜上成像的。实验的设计可以保持简单，因为我们发现，沙漠蚂蚁对真实全景的方向性反应可以通过粗略复制全景的天际线来引发。它暗示了蚂蚁可能用来识别场景的特征，比如天际线上的波峰和波谷。此外，由于距离和指南针信息对于场景识别并不是必不可少的，因此在实验室中研究它的各个方面是现实的。此外，暴露在大型视觉刺激下的幼稚蚂蚁往往面对其质量中心这一古老但长期被忽视的结果为我们提供了一种新的方法来分析蚂蚁如何在大场景中整合信息。我们的主要研究问题是：蚂蚁在识别过程中如何扫描和查看模式？蚂蚁和蜜蜂通过采用刻板印象的观察策略简化了模式识别的计算问题。我们将在蚂蚁检查、识别和使用不同大小的全景视觉场景时分析这些策略。飞行员数据显示，田野中的沙漠蚂蚁在修正路线之前会旋转扫描场景，或者如果场景不熟悉。这次扫描的视频将揭示全景中观看方向和引导元素之间的关系。在实验室中，将检查场景识别过程中观看行为的固定性或可变性以及场景转换引起的变化，以寻找场景编码坐标系(视网膜、指南针或场景固有的)的线索。蚂蚁如何在大范围内整合信息？作为一种分析大型场景中视觉处理的新方法，我们将研究蚂蚁对刺激质心的计算。因为这种瞄准行为不需要训练，所以我们可以测试各种刺激，并分析场景的不同区域和特征如何结合在一起来估计场景的质心。我们的目标是建立一个这一过程的模型，可以预测蚂蚁对大多数场景的方向性反应。全景场景的组成部分是如何编码的？通过模拟的天际线可以识别全景，这意味着场景的很大一部分被天际线轮廓捕捉到了。天际线编码的一种可能形式是定向边缘序列。另一种可能性是，场景组件的形状-峰值或低谷-通过天际线拐点和组件的质心之间的分离进行编码。通过训练蚂蚁辨别形状(最初是等腰三角形和等腰三角形)，我们将调查蚂蚁是否使用这些或其他视觉‘基元’。蚂蚁如何使用全景场景？我们将在现场研究来自自然全景图的信息(1)为导航提供方向信息，(2)作为上下文线索为被训练到两条觅食路线的蚂蚁启动适当的空间记忆，以及(3)在场景识别中与指南针信息相互作用的方式。

Panoramic scenes are important to many navigating animals in providing directional signals (home is left of the hill), compass direction (valley runs north-south), and a rough sense of place (we are nearly home). But how large natural scenes are encoded or recognised is almost unknown. We will explore this topic in ants for its intrinsic interest, as a new approach to studying insect pattern vision, and more broadly because the way that a low resolution and relatively simple visual system deals with natural scenes relates to current areas of neuroscience, computer vision and robotics. Walking ants are good insects for such a study as their scanning and viewing movements while recognising scenes and approaching patterns can be monitored at high resolution to give both the orientation of an ant's body axis and its position, allowing us to infer how a scene is imaged on the retina. The design of experiments can be kept simple following our finding that the directional responses of desert ants to a real panorama can be elicited by a crude facsimile of the skyline of that panorama. It hints at possible features that ants use for scene identification, such as peaks and troughs in the skyline. Also, as distance and compass information are not essential for scene recognition, it is realistic to study aspects of it in the lab. Additionally, the old but long-ignored result that naïve ants exposed to a large visual stimulus tend to face its centre of mass gives us a new way to analyse how ants integrate information across a large scene. Our major research questions are: How do ants scan and view patterns during recognition? Ants and bees simplify the computational problems of pattern recognition by adopting stereotyped viewing strategies. We will analyse these strategies while ants inspect, recognise and use panoramic visual scenes of different sizes. Pilot data show that desert ants in the field rotate to scan a scene before correcting their course, or if the scene is unfamiliar. Videos of this scanning will reveal the relationship between viewing direction and guiding elements in the panorama. In the lab, the fixity or variability of viewing behaviour during scene recognition and the changes induced by scene transformations will be examined for clues to the coordinate system of scene encoding (retinotopic, compass-based, or intrinsic to the scene). How do ants integrate information over a large area? As a new approach to analyzing the visual processing that occurs across large scenes, we will study the ants' computation of the centre of mass of a stimulus. Because this aiming behaviour needs no training, we can test a large variety of stimuli and analyse how different areas and features of a scene combine in the estimate of a scene's centre of mass. Our aim is to produce a model of this processing that can predict an ant's directional response to most scenes. How are components of panoramic scenes encoded? That a panorama can be recognised through its simulated skyline means that a significant part of the scene is captured by the skyline contour. One likely form of skyline encoding is as a sequence of oriented edges. Another possibility is that the shape of scene components - peaks or troughs - is encoded by the separation between skyline inflections and the centre of mass of a component. By training ants to distinguish between shapes (initially isosceles and scalene triangles), we will investigate whether ants use these or other visual 'primitives'. How do ants use panoramic scenes? We will examine in the field the ways that information from natural panoramas (1) provides directional information for guidance, (2) acts as a contextual cue for priming the appropriate spatial memories for ants that are trained to two foraging routes, and (3) interacts with compass information in scene recognition.