DSpace Repository :: Browsing by Author "Saavedra, Jose M."

Browsing by Author "Saavedra, Jose M."

Now showing 1 - 10 of 10

33
(33, 33) Pavez, Lukas; Saavedra, Jose M.; 33
During the last years, we have seen significant advances in the object detection task, mainly due to the outperforming results of convolutional neural networks. In this vein, anchor-based models have achieved the best results. However, these models require prior information about the aspect and scales of target objects, needing more hyperparameters to fit. In addition, using anchors to fit bounding boxes seems far from how our visual system does the same visual task. Instead, our visual system uses the interactions of different scene parts to semantically identify objects, called perceptual grouping. An object detection methodology closer to the natural model is anchor-free detection, where models like FCOS or Centernet have shown competitive results, but these have not yet exploited the concept of perceptual grouping. Therefore, to increase the effectiveness of anchor-free models keeping the inference time low, we propose to add non-local attention (NL modules) modules to boost the feature map of the underlying backbone. NL modules implement the perceptual grouping mechanism, allowing receptive fields to cooperate in visual representation learning. We show that non-local modules combined with an FCOS head (NL-FCOS) are practical and efficient. Thus, we establish state-of-the-art performance in clothing detection and handwritten amount recognition problems.
33
(33, 33) Saavedra, Jose M.; Morales, Javier; Murrugarra-Llerena, Nils; 33
Sketch-based image retrieval is demanding interest in the computer vision community due to its relevance in the visual perception system and its potential application in a wide diversity of industries. In the literature, we observe significant advances when the models are evaluated in public datasets. However, when assessed in real environments, the performance drops drastically. The big problem is that the SOTA SBIR models follow a supervised regimen, strongly depending on a considerable amount of labeled sketch-photo pairs, which is unfeasible in real contexts. Therefore, we propose SBIR-BYOL, an extension of the well-known BYOL, to work in a bimodal scenario for sketch-based image retrieval. To this end, we also propose a two-stage self-supervised training methodology, exploiting existing sketch-photo pairs and contour-photo pairs generated from photographs of a target catalog. We demonstrate the benefits of our model for the eCommerce environments, where searching is a critical component. Here, our self-supervised SBIR model shows an increase of over 60 % of mAP.
33
(33, 33) de la Sotta, Tomás; Chang, Violeta; Pizarro, Benjamín; Henriquez, Héctor; Alvear, Nicolás; Saavedra, Jose M.; 33
Chest x-ray images are one of the most commonly performed imaging tests, providing crucial clinical information about structures like the heart, lungs, ribs, bones, and blood vessels. In this context, image segmentation is a critical stage as it aims to separate significant parts of an image. However, manual image segmentation presents serious difficulties that can be tackled using deep learning-based methods, such as U-Net. Furthermore, attention mechanisms have attracted the interest of the machine-learning community and can bring improvements in medical image segmentation. This research aims to assess the impact of attention-based mechanisms for segmenting different organs in chest X-ray images, like heart, lungs, clavicles and ribs, over the well-known U-Net architecture as a baseline. We study five U-Net encoder variations, replacing the U-Net encoder with ResNet 18, 34 and 50, Swin Transformer and a simple residual structure comprised of a single ResNet-50 prior to each U-Net encoder layer. In the original U-Net, the skip layers are identity layers connecting each encoder block with its corresponding decoder one. Here, we replace these layers with different attention mechanisms: spatial attention, full spatial attention, double spatial attention, multiple spatial attention, spatial cross-attention and Swin spatial cross-attention. Each encoder variation and attention mechanism was evaluated from scratch and on a pre-trained scenario, independently for lungs, ribs, heart and clavicles. ResNet-UNet-18 achieves up to 0.96 and 0.53 of average overlapping with hand-segmented masks of lungs and clavicles, respectively. The best encoder for rib and heart segmentation was Residual U-Net, with 0.88 and 0.83 overlapping, respectively. Furthermore, for attention mechanisms, the most suitable were selected according to overlapping with hand-segmented masks as Spatial Attention U-Net for lungs (0.96), Three-Head Attention for ribs (0.88), Full Spatial Attention for the heart (0.82) and Spatial Cross-Attention for clavicles (0.54). Encoder variations and attention mechanism have a positive impact over a U-Net for segmentation of lungs, ribs, heart and clavicles from scratch, without transfer learning. Moreover, the encoder and the attention mechanism are not universal for segmenting different organs in chest X-ray images. While organs share the same backbone architecture (lungs and clavicles), the most appropriate attention mechanisms for each organ are all different, achieving up to 0.99 of overlapping in lung segmentation.
33
(33, 33) Arevalo-Ramirez, Tito; Alfaro, Anali; Saavedra, Jose M.; Recabarren, Matias; Ponce-Donoso, Mauricio; Delpiano, Jose; 33
Deep learning has gained popularity in recent years for reconstructing hyperspectral and multispectral images, offering cost-effective solutions and promising results. Research on hyperspectral image reconstruction feeds deep learning models with images at specific wavelengths and outputs images in other spectral bands. Although encouraging results of previous works, it should be determined to what extent the reconstructed information can lead to an advantage over the captured images. In this context, the present work inspects whether or not reconstructed spectral images add relevant information to segmentation networks for improving urban tree identification. Specifically, we generate red-edge (ReD) and near-infrared (NIR) images from RGB images using a conditional Generative Adversarial Network (cGAN). The training and validation are carried out with 5770 multispectral images obtained after a custom data augmentation process using an urban hyperspectral dataset. The testing outcomes reveal that ReD and NIR can be generated with an average Structural Similarity Index Measure of 0.93 and 0.88, respectively. Next, the cGAN generates ReD and NIR information of two RGB-based urban tree datasets (i.e., Jekyll, 3949 samples, and Arbocensus, 317 samples). Subsequently, DeepLabV3 and SegFormer segmentation networks are trained, validated, and tested using RGB, RGB+ReD, and RGB+NIR images from Jekyll and Arbocensus datasets. The experiments show that reconstructed multispectral images might not add information to segmentation networks that enhance their performance. Specifically, the p-values from a T-test show no significant difference between the performance of segmentation networks.
33
(33, 33) Donoso, Diego; Saavedra, Jose M.; 33
Sketch-based understanding is involved in human communication and cognitive development, making it essential in visual perception. A specific task in this domain is sketch-to-photo translation, where a model produces realistic images from simple drawings. To this end, large paired training datasets are commonly required, which is impractical in real applications. Thus, this work studies conditional generative models for sketch-to-photo translation, overcoming the lack of training datasets by a self-supervised approach that produces sketch-photo pairs from a target catalog. Our study shows the benefit of cycle-consistency loss and UNet architectures that, together with the proposed dataset generation, improve performance in real applications like eCommerce. Our results also reveal the weakness of conditional DDPMs for generating images resembling the input sketch, even though they achieve a high FID score.
33
(33, 33) Morales, Javier; Murrugarra-Llerena, Nils; Saavedra, Jose M.; 33
Sketch-based understanding is a critical component of human cognitive learning and is a primitive communication means between humans. This topic has recently attracted the interest of the computer vision community as sketching represents a powerful tool to express static objects and dynamic scenes. Unfortunately, despite its broad application domains, the current sketch-based models strongly rely on labels for supervised training, ignoring knowledge from unlabeled data, thus limiting the underlying generalization and the applicability. Therefore, we present a study about the use of unlabeled data to improve a sketch-based model. To this end, we evaluate variations of VAE and semi-supervised VAE, and present an extension of BYOL to deal with sketches. Our results show the superiority of sketch-BYOL, which outperforms other self-supervised approaches increasing the retrieval performance for known and unknown categories. Furthermore, we show how other tasks can benefit from our proposal.
Achieving high performance on sketch-based image retrieval without real sketches for training
Saavedra, Jose M.; Stears, Christopher; Campos, Waldo
Automatic floor plan analysis and recognition
Pizarro, Pablo N.; Hitschfeld, Nancy; Sipiran, Ivan; Saavedra, Jose M.
Challenges for computer vision as a tool for screening urban trees through street-view images
Arevalo-Ramirez, Tito; Alfaro, Anali; Figueroa, José; Ponce-Donoso, Mauricio; Saavedra, Jose M.; Recabarren, Matías; Delpiano, José
VETE: improving visual embeddings through text descriptions for eCommerce search engines
Martínez, Guillermo; Saavedra, Jose M.; Murrugara-Llerena, Nils

Browsing by Author "Saavedra, Jose M."

Results Per Page

Sort Options