Deep convolutional neural network (CNN) achieves outstanding performance in the field of target detection. As one of the most typical targets in remote sensing images (RSIs), airport has attracted increasing attention in recent years. However, the essential challenge for using deep CNN to detect airport is the great imbalance between the number of airports and background examples in large-scale RSIs, which may lead to over-fitting. In this paper, we develop a hard example mining and weight-balanced strategy to construct a novel end-to-end convolutional neural network for airport detection. The initial motivation of the proposed method is that backgrounds contain an overwhelming number of easy examples and a few hard examples. Therefore, we design a hard example mining layer to automatically select hard examples by their losses, and implement a new weight-balanced loss function to optimize CNN. Meanwhile, the cascade design of proposal extraction and object detection in our network releases the constraint on input image size and reduces spurious false positives. Compared with geometric characteristics and low-level manually designed features, the hard example mining based network could extract high-level features, which is more robust for airport detection in complex environment. The proposed method is validated on a multi-scale dataset with complex background collected from Google Earth. The experimental results demonstrate that our proposed method is robust, and superior to the state-of-the-art airport detection models.
9 May 2018 in IEEE Geoscience and Remote Sensing Letters
Convolutional neural network obtains remarkable achievements on target detection, due to its prominent capability on feature extraction. However, it still needs further study for aircraft detection task, since intraclass variation still restricts the accuracy of aircraft detection in remote sensing images. In this letter, we adopt regularity of aircraft circle response to design our end-to-end fully convolutional network (FCN), and embed online exemplar mining into our network to handle intraclass variation. The mined exemplars are employed to capture different intraclass characteristics, which effectively reduces the burden of network training. Specifically, we first select basic exemplars based on labeled information and initialize the relationships between exemplars and aircraft examples. Then, these relationships will be updated by the similarity of these examples in high-level features space. Finally, aircraft examples will be used to train different exemplar detectors according to updated relationships. Motivated by the geometric shape of aircraft, a circle response map is developed to construct our FCN to achieve more efficient aircraft detection. The comparative experiments indicate that superior performance of our network in accurate and efficient aircraft detection.
27 March 2024 in IEEE Transactions on Pattern Analysis and Machine Intelligence
This paper studies how to flexibly integrate reconstructed 3D models into practical 3D modeling pipelines such as 3D scene creation and rendering. Due to the technical difficulty, one can only obtain rough 3D models (R3DMs) for most real objects using existing 3D reconstruction techniques. As a result, physically-based rendering (PBR) would render low-quality images or videos for scenes that are constructed by R3DMs. One promising solution would be representing real-world objects as Neural Fields such as NeRFs, which are able to generate photo-realistic renderings of an object under desired viewpoints. However, a drawback is that the synthesized views through Neural Fields Rendering (NFR) cannot reflect the simulated lighting details on R3DMs in PBR pipelines, especially when object interactions in the 3D scene creation cause local shadows. To solve this dilemma, we propose a lighting transfer network (LighTNet) to bridge NFR and PBR, such that they can benefit from each other. LighTNet reasons about a simplified image composition model, remedies the uneven surface issue caused by R3DMs, and is empowered by several perceptual-motivated constraints and a new Lab angle loss which enhances the contrast between lighting strength and colors. Comparisons demonstrate that LighTNet is superior in synthesizing impressive lighting, and is promising in pushing NFR further in practical 3D modeling workflows.
Neural Radiance Fields (NeRF) have emerged as a potent paradigm for representing scenes and synthesizing photo-realistic images. A main limitation of conventional NeRFs is that they often fail to produce high-quality renderings under novel viewpoints that are significantly different from the training viewpoints. In this paper, instead of exploiting few-shot image synthesis, we study the novel view extrapolation setting that (1) the training images can well describe an object, and (2) there is a notable discrepancy between the training and test viewpoints' distributions. We present RapNeRF (RAy Priors) as a solution. Our insight is that the inherent appearances of a 3D surface's arbitrary visible projections should be consistent. We thus propose a random ray casting policy that allows training unseen views using seen views. Furthermore, we show that a ray atlas pre-computed from the observed rays' viewing directions could further enhance the rendering quality for extrapolated views. A main limitation is that RapNeRF would remove the strong view-dependent effects because it leverages the multi-view consistency property.