Locate Globally, Segment Locally: A Progressive Architecture With Knowledge Review Network for Salient Object Detection

Abstract

Salient object location and segmentation are two different tasks in salient object detection (SOD). The former aims to globally find the most attractive objects in an image, whereas the latter can be achieved only using local regions that contain salient objects. However, previous methods mainly accomplish the two tasks simultaneously in a simple end-to-end manner, which leads to the ignorance of the differences between them. We assume that the human vision system orderly locates and segments objects, so we propose a novel progressive architecture with knowledge review network (PA-KRN) for SOD. It consists of three parts. (1) A coarse locating module (CLM) that uses body-attention label locates rough areas containing salient objects without boundary details. (2) An attention-based sampler highlights salient object regions with high resolution based on body-attention maps. (3) A fine segmenting module (FSM) finely segments salient objects. The networks applied in CLM and FSM are mainly based on our proposed knowledge review network (KRN) that utilizes the finest feature maps to reintegrate all previous layers, which can make up for the important information that is continuously diluted in the top-down path. Experiments on five benchmarks demonstrate that our single KRN can outperform state-of-the-art methods. Furthermore, our PA-KRN performs better and substantially surpasses the aforementioned methods.

View-Aware Salient Object Detection for $360^{\circ }$ Omnidirectional Image

“…To verify the challenges of the proposed dataset ODI-SOD, in Tab.II we list the performances of 20 state-ofthe-art (SOTA) 2D SOD and 360 • SOD methods on our test set without finetuning on our train set. The methods include GCPANet [46], MINet-R [45], ITSD [69], F3Net [70], DFI [71], PFSNet [72], CTDNet [51], VST [50], PAKRN [9], DCN [73], SOD100K [74], PSGLoss [75], SCASOD [76], FastSaliency [77], PurNet [78], PoolNet [79], RCSB [80], ZoomNet [10], TRACER [81] and DDS [7]. From Tab.II we find that all listed methods perform not well on the ODI-SOD test set including the 360 • -based method DDS [7] trained by the 360-SOD train set, which suggests that currently available models have poor generalization ability over the proposed dataset.…”

“…In contrast, the severely distorted regions can easily lead to segmentation failure due to their apparent differences from existing knowledge and perception. The target objects with discontinuous edge effects and the To demonstrate the effectiveness of the proposed method, we selected the Top-5 in performance from 2D SOD methods with available training code in Tab.II, i.e., PAKRN [9], PoolNet [79], RCSB [80], ZoomNet [10] and TRACER [81]. Then, for a fair comparison, we finetune these five models, DDS [7] and our proposed model on the ODI-SOD train set.…”

“…After finetuning, their evaluation results on the ODI-SOD test set are listed in Tab.III. We can see that the overall metric scores are much better than those before finetuning in Tab.II for the selected methods except for PAKRN [9]. One primary reason may be that PAKRN [9] needs multi-stage joint training, which is not easy for a new task.…”

“…Moreover, through the observation of the dataset, we find that the poor performances of existing state-of-the-art methods [7], [9], [10] can attribute to three prominent challenges, i.e., diverse distortion degrees, discontinuous edge effects and changeable object scales. Distortion varying with the projection position (e.g., the first col in Fig.…”

MPTC-FPN: A Multilayer Progressive FPN With Transformer-CNN Based Encoder for Salient Object Detection

“…We compare our proposed method with 16 previous stateof-the-art methods, including Amulet [55], UCF [56], BRN [45], C2SNet [23], AFNet [11], BASNet [37], F3Net [48], CAGNet-R [34], GCPANet [7], ITSD [60], LDF [49], MINet [35], GateNet [59], VST [28], PAKRN [51], PFSNet [31]. To ensure the fairness of the comparison, all predicted saliency maps are downloaded from the public official website and evaluated under the same evaluation code and environment.…”