TJU-DHD: A Diverse High-Resolution Dataset for Object Detection

Pang, Yanwei; Cao, Jiale; Li, Yazhao; Xie, Jin; Sun, Hanqing; Gong, Jinfeng

doi:10.1109/tip.2020.3034487

Cited by 67 publications

(27 citation statements)

References 87 publications

(197 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It is straightforward to replace it by a one-pass deep object detector to speed up tracking, with the expense of lower accuracy. Moreover, it is possible to improve the tracking performance by pre-training the detector on a high resolution dataset [66].…”

Section: Discussionmentioning

confidence: 99%

TDIOT: Target-driven Inference for Deep Video Object Tracking

Gurkan,

Cerkezi,

Cirakman

et al. 2021

Preprint

View full text Add to dashboard Cite

Recent tracking-by-detection approaches use deep object detectors as target detection baseline, because of their high performance on still images. For effective video object tracking, object detection is integrated with a data association step performed by either a custom design inference architecture or an end-to-end joint training for tracking purpose. In this work, we adopt the former approach and use the pre-trained Mask R-CNN deep object detector as the baseline. We introduce a novel inference architecture placed on top of FPN-ResNet101 backbone of Mask R-CNN to jointly perform detection and tracking, without requiring additional training for tracking purpose. The proposed single object tracker, TDIOT, applies an appearance similarity-based temporal matching for data association. In order to tackle tracking discontinuities, we incorporate a local search and matching module into the inference head layer that exploits SiamFC for short term tracking. Moreover, in order to improve robustness to scale changes, we introduce a scale adaptive region proposal network that enables to search the target at an adaptively enlarged spatial neighborhood specified by the trace of the target. In order to meet long term tracking requirements, a low cost verification layer is incorporated into the inference architecture to monitor presence of the target based on its LBP histogram model. Performance evaluation on videos from VOT2016, VOT2018 and VOT-LT2018 datasets demonstrate that TDIOT achieves higher accuracy compared to the state-of-the-art short-term trackers while it provides comparable performance in long term tracking.

show abstract

Section: Discussionmentioning

confidence: 99%

TDIOT: Target-driven Inference for Deep Video Object Tracking

Gurkan,

Cerkezi,

Cirakman

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…On NightOwls dataset, following [15], [53] and the official NightOwls evaluation application programming interface (API) 1 , we report the evaluation results on the Reasonable subset (non-occluded pedestrians with height >= 50 pixels), Reasonable small (non-occluded pedestrians with height between 50 pixels and 75 pixels), Reasonable occ subsets (occluded pedestrians with height >= 50 pixels), and All (pedestrians with height >= 20 pixels). On TJU-DHDpedestrian dataset, we report evaluation results on R, HO, R+HO and All subsets, which is the same with Pang et al [54].…”

Section: A Experimental Setup 1) Datasetsmentioning

confidence: 98%

“…We evaluate our proposed method on three standard benchmark datasets including CityPersons [9], NightOwls [53] and TJU-DHD-pedestrian [54] involving both daytime and nighttime pedestrian detection. CityPersons is collected for daytime pedestrian detection.…”

Section: A Experimental Setup 1) Datasetsmentioning

confidence: 99%

See 1 more Smart Citation

Pedestrian Detection by Exemplar-Guided Contrastive Learning

Lin¹,

Pei²,

Chen³

et al. 2021

Preprint

View full text Add to dashboard Cite

“…Existing state-of-the-art (SOTA) deep learning-based object detection models are trained and evaluated on large-scale datasets such as ImageNet [21], Pascal VOC [22], and MS COCO [23] which are generic and largely represent medium and large scale objects. Although existing SOTA deep learning-based object detection models perform well for medium and large scale object detection tasks their direct application for specific object detection tasks, especially for small objects like malaria parasite screening, will not achieve a very good detection performance [24] [25]. When an input image passes through different layers of SOTA detectors it will lose too much spatial information which is crucial for small object localization.…”

Section: Problem Statement and Motivationmentioning

confidence: 99%

Tile-Based Microscopic Image Processing For Malaria Screening Using Deep Learning Approach

Abdurahman

Fante

2022

Preprint

View full text Add to dashboard Cite

Background: Manual microscopic examination remains the golden standard for malaria diagnosis. However, it is cumbersome and requires the experience of pathologists for accurate diagnosis. The shear workload and challenges involved in manual microscopy drives for alternative computer-aided diagnosis techniques. While the importance of computer-aided diagnosis is increasing at an enormous pace fostered by the advancement of deep learning algorithms, there are still challenges. Existing state-of-the-art (SOTA) deep learning-based object detection models suffer from effectively detecting small objects which are less represented on benchmark datasets and are affected by the loss of detailed spatial information due to in-network feature map downscaling. The problem even becomes harder when the input image is high resolution. This is due to the fact that existing SOTA models will not directly process high-resolution images limited by their low-resolution network input size. Methods: In this study, an effective and robust tile-based image processing approach was proposed to enhance malaria parasite detection performance of SOTA object detection models. Three YOLOV4 based object detectors were adopted considering their detection speed and accuracy. The proposed detection models were developed by using tiles generated from 1,780 high-resolution P. falciparum-infected thick smear microscopic images. External validation was performed to verify the detection accuracy and generalization ability of the proposed models on three datasets acquired from a different domain.Results: The best-performing model leveraging the proposed tile-based approach significantly outperforms its baseline method (Recall, [95.3%] vs [57%] and Average Precision, [87.1%] vs [76%]). In addition, the proposed method achieved state-of-the-art performance compared with previous research works which use different machine learning techniques on similar datasets.Conclusions: The obtained qualitative and empirical results show that the proposed method reveals a fundamental performance improvement for the detection of P. falciparum from thick smear microscopic images while maintaining real-time speed. Furthermore, the proposed method could have the potential to assist and reduce the workload of laboratory technicians in malaria-endemic remote areas in developing countries where there is a critical skill gap and shortage of experts.

show abstract

TJU-DHD: A Diverse High-Resolution Dataset for Object Detection

Cited by 67 publications

References 87 publications

TDIOT: Target-driven Inference for Deep Video Object Tracking

TDIOT: Target-driven Inference for Deep Video Object Tracking

Pedestrian Detection by Exemplar-Guided Contrastive Learning

Tile-Based Microscopic Image Processing For Malaria Screening Using Deep Learning Approach

Contact Info

Product

Resources

About