As capturing devices become common, 3D scans of interior spaces are acquired on a daily basis. Through scene comparison over time, information about objects in the scene and their changes is inferred. This information is important for robots and AR and VR devices, in order to operate in an immersive virtual experience. We thus propose an unsupervised object discovery method that identifies added, moved, or removed objects without any prior knowledge of what objects exist in the scene. We model this problem as a combination of a 3D change detection and a 2D segmentation task. Our algorithm leverages generic 2D segmentation masks to refine an initial but incomplete set of 3D change detections. The initial changes, acquired through render-and-compare likely correspond to movable objects. The incomplete detections are refined through graph optimization, distilling the information of the 2D segmentation masks in the 3D space. Experiments on the 3Rscan dataset prove that our method outperforms competitive baselines, with SoTA results.
Unsupervised object-centric learning aims to decompose scenes into interpretable object entities, termed slots. Slot-based auto-encoders stand out as a prominent method for this task. Within them, crucial aspects include guiding the encoder to generate object-specific slots and ensuring the decoder utilizes them during reconstruction. This work introduces two novel techniques, (i) an attention-based self-training approach, which distills superior slot-based attention masks from the decoder to the encoder, enhancing object segmentation, and (ii) an innovative patch-order permutation strategy for autoregressive transformers that strengthens the role of slot vectors in reconstruction. The effectiveness of these strategies is showcased experimentally. The combined approach significantly surpasses prior slot-based autoencoder methods in unsupervised object segmentation, especially with complex real-world images. We provide the implementation code at https://github.com/gkakogeorgiou/spot .
Over the last decade there has been an increasing frequency and intensity of wildfires across the globe, posing significant threats to human and animal lives, ecosystems, and socio-economic stability. Therefore urgent action is required to mitigate their devastating impact and safeguard Earth's natural resources. Robust Machine Learning methods combined with the abundance of high-resolution satellite imagery can provide accurate and timely mappings of the affected area in order to assess the scale of the event, identify the impacted assets and prioritize and allocate resources effectively for the proper restoration of the damaged region. In this work, we create and introduce a machine-learning ready dataset we name FLOGA (Forest wiLdfire Observations for the Greek Area). This dataset is unique as it comprises of satellite imagery acquired before and after a wildfire event, it contains information from Sentinel-2 and MODIS modalities with variable spatial and spectral resolution, and contains a large number of events where the corresponding burnt area ground truth has been annotated by domain experts. FLOGA covers the wider region of Greece, which is characterized by a Mediterranean landscape and climatic conditions. We use FLOGA to provide a thorough comparison of multiple Machine Learning and Deep Learning algorithms for the automatic extraction of burnt areas, approached as a change detection task. We also compare the results to those obtained using standard specialized spectral indices for burnt area mapping. Finally, we propose a novel Deep Learning model, namely BAM-CD. Our benchmark results demonstrate the efficacy of the proposed technique in the automatic extraction of burnt areas, outperforming all other methods in terms of accuracy and robustness. Our dataset and code are publicly available at: https://github.com/Orion-AI-Lab/FLOGA.
Convolutional networks and vision transformers have different forms of pairwise interactions, pooling across layers and pooling at the end of the network. Does the latter really need to be different? As a by-product of pooling, vision transformers provide spatial attention for free, but this is most often of low quality unless self-supervised, which is not well studied. Is supervision really the problem? In this work, we develop a generic pooling framework and then we formulate a number of existing methods as instantiations. By discussing the properties of each group of methods, we derive SimPool, a simple attention-based pooling mechanism as a replacement of the default one for both convolutional and transformer encoders. We find that, whether supervised or self-supervised, this improves performance on pre-training and downstream tasks and provides attention maps delineating object boundaries in all cases. One could thus call SimPool universal. To our knowledge, we are the first to obtain attention maps in supervised transformers of at least as good quality as self-supervised, without explicit losses or modifying the architecture. Code at: https://github.com/billpsomas/simpool.
Mapping the seafloor with underwater imaging cameras is of significant importance for various applications including marine engineering, geology, geomorphology, archaeology and biology. For shallow waters, among the underwater imaging challenges, caustics i.e., the complex physical phenomena resulting from the projection of light rays being refracted by the wavy surface, is likely the most crucial one. Caustics is the main factor during underwater imaging campaigns that massively degrade image quality and affect severely any 2D mosaicking or 3D reconstruction of the seabed. In this work, we propose a novel method for correcting the radiometric effects of caustics on shallow underwater imagery. Contrary to the state-of-the-art, the developed method can handle seabed and riverbed of any anaglyph, correcting the images using real pixel information, thus, improving image matching and 3D reconstruction processes. In particular, the developed method employs deep learning architectures in order to classify image pixels to "non-caustics" and "caustics". Then, exploits the 3D geometry of the scene to achieve a pixel-wise correction, by transferring appropriate color values between the overlapping underwater images. Moreover, to fill the current gap, we have collected, annotated and structured a real-world caustic dataset, namely R-CAUSTIC, which is openly available. Overall, based on the experimental results and validation the developed methodology is quite promising in both detecting caustics and reconstructing their intensity.
AR/VR applications and robots need to know when the scene has changed. An example is when objects are moved, added, or removed from the scene. We propose a 3D object discovery method that is based only on scene changes. Our method does not need to encode any assumptions about what is an object, but rather discovers objects by exploiting their coherent move. Changes are initially detected as differences in the depth maps and segmented as objects if they undergo rigid motions. A graph cut optimization propagates the changing labels to geometrically consistent regions. Experiments show that our method achieves state-of-the-art performance on the 3RScan dataset against competitive baselines. The source code of our method can be found at https://github.com/katadam/ObjectsCanMove.
Data association is a crucial component for any multiple object tracking (MOT) method that follows the tracking-by-detection paradigm. To generate complete trajectories such methods employ a data association process to establish assignments between detections and existing targets during each timestep. Recent data association approaches try to solve a multi-dimensional linear assignment task or a network flow minimization problem or either tackle it via multiple hypotheses tracking. However, during inference an optimization step that computes optimal assignments is required for every sequence frame adding significant computational complexity in any given solution. To this end, in the context of this work we introduce Transformer-based Assignment Decision Network (TADN) that tackles data association without the need of any explicit optimization during inference. In particular, TADN can directly infer assignment pairs between detections and active targets in a single forward pass of the network. We have integrated TADN in a rather simple MOT framework, we designed a novel training strategy for efficient end-to-end training and demonstrate the high potential of our approach for online visual tracking-by-detection MOT on two popular benchmarks, i.e. MOT17 and UA-DETRAC. Our proposed approach outperforms the state-of-the-art in most evaluation metrics despite its simple nature as a tracker which lacks significant auxiliary components such as occlusion handling or re-identification. The implementation of our method is publicly available at https://github.com/psaltaath/tadn-mot.
Transformers and masked language modeling are quickly being adopted and explored in computer vision as vision transformers and masked image modeling (MIM). In this work, we argue that image token masking is fundamentally different from token masking in text, due to the amount and correlation of tokens in an image. In particular, to generate a challenging pretext task for MIM, we advocate a shift from random masking to informed masking. We develop and exhibit this idea in the context of distillation-based MIM, where a teacher transformer encoder generates an attention map, which we use to guide masking for the student encoder. We thus introduce a novel masking strategy, called attention-guided masking (AttMask), and we demonstrate its effectiveness over random masking for dense distillation-based MIM as well as plain distillation-based self-supervised learning on classification tokens. We confirm that AttMask accelerates the learning process and improves the performance on a variety of downstream tasks.
Metric learning involves learning a discriminative representation such that embeddings of similar classes are encouraged to be close, while embeddings of dissimilar classes are pushed far apart. State-of-the-art methods focus mostly on sophisticated loss functions or mining strategies. On the one hand, metric learning losses consider two or more examples at a time. On the other hand, modern data augmentation methods for classification consider two or more examples at a time. The combination of the two ideas is under-studied. In this work, we aim to bridge this gap and improve representations using mixup, which is a powerful data augmentation approach interpolating two or more examples and corresponding target labels at a time. This task is challenging because, unlike classification, the loss functions used in metric learning are not additive over examples, so the idea of interpolating target labels is not straightforward. To the best of our knowledge, we are the first to investigate mixing examples and target labels for deep metric learning. We develop a generalized formulation that encompasses existing metric learning loss functions and modify it to accommodate for mixup, introducing Metric Mix, or Metrix. We show that mixing inputs, intermediate representations or embeddings along with target labels significantly improves representations and outperforms state-of-the-art metric learning methods on four benchmark datasets.
Although deep neural networks hold the state-of-the-art in several remote sensing tasks, their black-box operation hinders the understanding of their decisions, concealing any bias and other shortcomings in datasets and model performance. To this end, we have applied explainable artificial intelligence (XAI) methods in remote sensing multi-label classification tasks towards producing human-interpretable explanations and improve transparency. In particular, we developed deep learning models with state-of-the-art performance in the benchmark BigEarthNet and SEN12MS datasets. Ten XAI methods were employed towards understanding and interpreting models' predictions, along with quantitative metrics to assess and compare their performance. Numerous experiments were performed to assess the overall performance of XAI methods for straightforward prediction cases, competing multiple labels, as well as misclassification cases. According to our findings, Occlusion, Grad-CAM and Lime were the most interpretable and reliable XAI methods. However, none delivers high-resolution outputs, while apart from Grad-CAM, both Lime and Occlusion are computationally expensive. We also highlight different aspects of XAI performance and elaborate with insights on black-box decisions in order to improve transparency, understand their behavior and reveal, as well, datasets' particularities.