The RailEye3D dataset, a collection of train-platform scenarios for applications targeting passenger safety and automation of train dispatching, consists of 10 image sequences captured at 6 railway stations in Austria. Annotations for multi-object tracking are provided in both an unified format as well as the ground-truth format used in the MOTChallenge.

RoboBEV is a robustness evaluation benchmark tailored for camera-based bird's eye view (BEV) perception under natural data corruptions and domain shift. It includes eight distinct corruptions, including Bright, Dark, Fog, Snow, Motion Blur, Color Quant, Camera Crash, and Frame Lost. Source: RoboBEV: Towards Robust Bird's Eye View Perception under Corruptions Image Source: RoboBEV: Towards Robust Bird's Eye View Perception under Corruptions

An open source Multi-View Overhead Imagery dataset with 27 unique looks from a broad range of viewing angles (-32.5 degrees to 54.0 degrees). Each of these images cover the same 665 square km geographic extent and are annotated with 126,747 building footprint labels, enabling direct assessment of the impact of viewpoint perturbation on model performance. Source: SpaceNet MVOI: a Multi-View Overhead Imagery Dataset

A new dataset for streaming classification consisting of temporally correlated images from 51 distinct object categories and additional evaluation classes outside of the training distribution to test novelty recognition. Source: Stream-51: Streaming Classification and Novelty Detection from Videos

We propose a large-scale underwater instance segmentation dataset, UIIS10K, which includes 10,048 images with pixel-level annotations for 10 categories. As far as we know, this is the largest underwater instance segmentation dataset available and can be used as a benchmark for evaluating underwater segmentation methods.

The newly introduced UP-COUNT dataset includes drone footage captured with cameras from the DJI Mini 2 family UAV. It encompasses diverse environments, including streets, plazas, public transport stops, parks and other green recreation places. We recorded 202 unique videos and then extracted frames with a step of one second, resulting in 10,000 images with a resolution of 3840 × 2160 pixels. The recordings were taken at different altitudes and speeds of flight, and with various densities of people. Acquisition conditions vary in daytime and lighting, creating challenging shadows. Extra altitude information is provided for each image. Next, the labels of people’s heads were hand-prepared, resulting in 352,487 instances. During the labelling process, each image was marked and checked by two different people, and the continuity of labels within each sequence was reviewed. The lowest- (26.0 meters) and the highest-altitude (101.0 meters) recorded among the sequences, with an average of 60.3 meters. The bottom image presents the most crowded image (1,039 people instances), while the average object count is 35.25. Increased variability in crowd counts and different backgrounds caused by the lack of a stationary camera position better reflects real-world scenarios. The UP-COUNT dataset is divided into three subsets for training, validation and testing purposes, containing 141, 30 and 31 sequences, respectively. The described sequences’ splits are prepared using altitude-based stratified sampling, providing a comparable altitude distribution between the dataset’s splits.

Vehicle-Rear is a novel dataset for vehicle identification that contains more than three hours of high-resolution videos, with accurate information about the make, model, color and year of nearly 3,000 vehicles, in addition to the position and identification of their license plates.

360-SOD contains 500 high-resolution equirectangular images. Source: Distortion-adaptive Salient Object Detection in 360$^\circ$ Omnidirectional Images Image Source: http://cvteam.net/projects/JSTSP20_DDS/DDS.html

Description： 5,011 Images – Human Frontal face Data (Male). The data diversity includes multiple scenes, multiple ages and multiple races. This dataset includes 2,004 Caucasians , 3,007 Asians. This dataset can be used for tasks such as face detection, race detection, age detection, beard category classification. Data size： 5,011 people, one image per person Race distribution： 2,004 Caucasians , 3,007 Asians

A real-world dataset, with hyper-accurate digital counterpart & comprehensive ground-truth annotation. Dataset Content - 200 sequences (~400 mins) - 398 objects (324 stationary, 74 dynamic) - 2 real indoor scenes - Single + multi-user activities Sensor Data per device - 2 x outward-facing monochrome camera streams - 1 x outward-facing RGB camera stream - 2 x IMU streams - 2 x Internal-facing eye tracking cameras - Complete sensor calibrations

Bangladeshi Sign Language Image Dataset (BdSLImset) is a dataset that contains images of different Bangladeshi sign letters. Source: https://github.com/imruljubair/BdSLImset Image Source: https://arxiv.org/pdf/1811.12813v1.pdf

C2A: Combination to Application Dataset ## Overview This repository contains the code and information for the paper "UAV-Enhanced Combination to Application: Comprehensive Analysis and Benchmarking of a Human Detection Dataset for Disaster Scenarios" by Ragib Amin Nihal, Benjamin Yen, Katsutoshi Itoyama, and Kazuhiro Nakadai. The C2A (Combination to Application) dataset is a novel synthetic dataset designed to advance human detection in disaster scenarios using UAV imagery. It combines disaster scene backgrounds from the AIDER dataset with human poses from the LSP/MPII-MPHB dataset to create a comprehensive resource for training machine learning models. The full paper is available at: https://arxiv.org/pdf/2408.04922 ## Dataset The C2A dataset consists of 10,215 images containing over 360,000 annotated human instances in various disaster scenarios. It includes diverse human poses (bent, kneeling, lying, sitting, upright) and disaster contexts (traffic incidents, fire, flood, collapsed buildings). ### Usage Notes - The dataset is split into training, validation, and test sets. - Two annotation formats are provided for flexibility: YOLO and COCO. - Pose information is available in a separate folder for all images. - Users can choose between standard object detection (YOLO/COCO) or pose-aware detection (All labels with Pose info). ## Key Features - Synthetic dataset combining real disaster scenes with human poses - Over 360,000 annotated human instances - 5 human pose categories - 4 disaster scenario types - Image resolutions ranging from 123x152 to 5184x3456 pixels - Designed to improve human detection in complex disaster environments

The Caltech Cars dataset consists of 126 rear-view photographs captured within parking lots. These images possess a resolution of 896 × 592 pixels, featuring a solitary vehicle as the primary subject. The acquisitions were made during daylight hours employing a handheld camera at roughly equivalent distances for all instances. "The dataset was presented in a PhD thesis titled 'Unsupervised learning of models for object recognition.'"

CholecTrack20 is a surgical video dataset focusing on laparoscopic cholecystectomy and designed for surgical tool tracking, featuring 20 annotated videos. The dataset includes detailed labels for multi-class multi-tool tracking, offering trajectories for tool visibility within the camera scope, intracorporeal movement within the patient's body, and the life-long intraoperative trajectory of each tool. Annotations cover spatial coordinates, tool class, operator identity, phase, visual conditions (occlusion, bleeding, smoke), and more for tools like grasper, bipolar, hook, scissors, clipper, irrigator, and specimen bag, with annotations provided at 1 frame per second across 35K frames and 65K instance tool labels. The dataset uses official splits, allocating 10 videos for training, 2 for validation, and 8 for testing. Chinedu Nwoye

The training and validation data are subsets of the training split of the MS COCO dataset (2017 release, bounding boxes only). The test set is taken from the validation split of the MS COCO dataset.

COMPASS-XP is a dataset of matched photographic and X-ray images of single objects, made available for use in Machine Learning & Computer Vision research, in particular in the context of transport security. Objects are imaged in multiple poses, and accompanied by metadata including labels for whether we consider the object to be dangerous in the context of aviation. Object classes overlap with those in the popular ImageNet Large Scale Visual Recognition Challenge class set and theWordNet lexical database, and identifiers for shared classes in both schemes are also provided. Hardware Configuration Photographs were captured with a Sony DSC-W800 compact digital camera. X-ray scans were obtained using a Gilardoni FEP ME 536 mailroom X-ray machine, distributed in the UK by Todd Research under the name TR50. The scanner is dual energy and generates several image outputs: • Low: Raw 8-bit greyscale data from the scanner’s low energy X-ray channel. • High: Raw 8-bit greyscale data from the scanner’s high energy X-ray channel. • Density: 8-bit greyscale data representing inferred material density computed from the two channels. • Grey: RGB PNG image representing a combination of both low and high energy channels with some appearance improvements. Although nominally greyscale, the image does include subtle duotone-style colouration. • Colour RGB PNG image with false colour palette representing material density. In practice the grey and colour versions are probably most useful, but for completeness the dataset includes all variants for each scan. Data Files Image files are supplied in six subdirectories, corresponding to the five X-ray image variants above plus photos. X-rays are provided in PNG format, while photos are JPEG. Each scan is identified by a numeric index, which is also used to name the file, padded with leading zeros to always be 4 digits long. Scan metadata is provided in the accompanying tab-delimited text file, meta.txt. This includes the following columns: • basename: The zero-padded identifier for the scan. All six image type variants for the same class-instance-pose have the same basename. X-ray files are named basename.png while photos are basename.jpg. • class: The object class in the scan. • instance: An integer identifying the object instance. Instances start at 1 for each class. • pose: An integer identifying the object pose. Poses start at 1 for each instance. • scan tray: Either A, indicating that the pose was imaged in a weighted tray, or N indicating it was not. • dangerous: Whether the object was considered dangerous (True/False). • IN id: Numeric index of the object class in the ILSVRC list of 1000 classes, or empty if the class isn’t present there. • WN id: WordNet identifier for the object class, or empty if the class isn’t present inWordNet. License The COMPASS-XP dataset was acquired as part of a research project funded by the UK Government Future Aviation Security Solutions programme. Both the images and their metadata are licensed under the Creative Commons Attribution 4.0 International License and may be freely used for research and commercial purpose, including derivative works, providing the source is acknowledged. COMPASS-XP Dataset Authors Lewis D. Griffin*, Matthew Caldwell, Jerone T. A. Andrews Computational Security Science Group, UCL * l.griffin@cs.ucl.ac.uk

CPPE - 5 (Medical Personal Protective Equipment) is a new challenging dataset with the goal to allow the study of subordinate categorization of medical personal protective equipments, which is not possible with other popular data sets that focus on broad level categories. Some features of this dataset are: * high quality images and annotations (~4.6 bounding boxes per image) * real-life images unlike any current such dataset * majority of non-iconic images (allowing easy deployment to real-world environments)

DeepPCB * Dataset Link : * A dataset contains 1,500 image pairs, each of which consists of a defect-free template image and an aligned tested image with annotations including positions of 6 most common types of PCB defects: open, short, mousebite, spur, pin hole, and spurious copper. ### Dataset Description #### Image Collection * All the images in this dataset are obtained from a linear scan CCD in a resolution of around 48 pixels per 1 millimeter. * The defect-free template images are manually checked and cleaned from sampled images in the above manner. * The original size of the template and the tested image is around 16k x 16k pixels. * Then they are cropped into many sub-images with the size of 640 x 640 and aligned through template matching techniques. * Next, a threshold is carefully selected to employ binarization to avoid illumination disturbance. * Notice that pre-processing algorithms can be varied according to the specific PCB defect detection algorithms, however, the image registration and thresholding techniques are common processes for high-accuracy PCB defect localization and classification. * An example pair in DeepPCB dataset is illustrated in the following figure, where the right one is the defect-free template image and the left one is the defective tested image with the ground truth annotations. <!--| An example of the tested image | The corresponding template image | |---|---| | ![tested image](https://github.com/tangsanli5201/DeepPCB/blob/08e98c4db5922613fb97176eb3d6497d48260cb1/fig/test.jpg") |  | --> #### Image Annotation We use the axis-aligned bounding box with a class ID for each defect in the tested images. As illustrated in above, we annotate six common types of PCB defects: open, short, mousebite, spur, pin hole, and spurious copper. Since there are only a few defects in the real tested image, we manually arguement some artificial defects on each tested image according to the PCB defect patterns, which leads to around 3 to 12 defects in each 640 x 640 image. The number of PCB defects is shown in the following figure. We separate 1,000 images as a training set and the remains as a test set. Each annotated image owns an annotation file with the same filename, e.g.00041000_test.jpg, 00041000_temp.jpg, and 00041000.txt are the tested image, template image, and the corresponding annotation file. Each defect on the tested image is annotated as the format:x1,y1,x2,y2, type, where (x1,y1) and (x2,y2) is the top left and the bottom right corner of the bounding box of the defect. type is an integer ID that follows the matches: 0-background (not used), 1-open, 2-short, 3-mousebite, 4-spur, 5-copper, 6-pin-hole. <!-- <div align=center> <img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/CountPCB.png" width="560"> </div> --> The annotation tool is now available with the source code in the ./tools directory. #### Benchmarks The average precision rate and F-score are used for evaluation. A detection is correct only if the intersection of unit (IoU) between the detected bounding box and any of the ground truth box with the same class is larger than 0.33. F-score is calculated as: F-score=2PR/(P+R), where P and R is the precision and recall rate. Notice that F-score is threshold-sensitive, which means you could adjust your score threshold to obtain a better result. Although F-score is not as fair as the mAP criteria but more practical since a threshold should always be given when deploying the model and not all of the algorithms have a score evaluation for the target. Thus, F-score and mAP are both under consideration in the benchmarks. The evaluation script for mAP and F-score are borrowed from Icdar2015 evaluation scripts with a small modification (You may first register an account.). Here, we give the modified evaluation scripts and the ground truth gt.zip file of the test set in evaluation/ directory. You can evaluate your own method by following the instructions: * run your algorithm and save the detected result for each image named as image_name.txt, where the image_name should be the same as in the gt.zip. You should follow the format of evaluation/gt.zip except that the output description of each defect from your algorithm should be: x1,y1,x2,y2, confidence, type, where (x1,y1) and (x2,y2) is the top left and the bottom right corner of the bounding box of the defect. confidence is a float number to show how confident you believe such detection results. type is a string and should be one of the following: open, short,mousebite,spur,copper,pin-hole. Notice there is no space except the comma. * zip your .txt file to res.zip. (You should not contain any sub-directory in the res.zip file) * run the evaluation script: python script.py -s=res.zip -g=gt.zip ### Approach This section with the source code will be public after the acceptance of the paper. #### Experiment results Here we show some results of our model based on deep neural networks. Our model achieves 98.6% mAp, 98.2% F-score @ 62FPS. More statistical analysis will be public after the acceptance of the paper. The green bounding box is the predicted location of the PCB defect with the confidence on the top of each. <!-- Result pair 1: <div align=center><img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_test1.jpg" width="375" style="margin:20"> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; <img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_temp1.jpg" width="375" style="margin:20"> </div> Result pair 2: <div align=center><img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_test2.jpg" width="375" style="margin:20"> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; <img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_temp2.jpg" width="375" style="margin:20"> </div> Result pair 3: <div align=center><img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_test3.jpg" width="375" style="margin:20"> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; <img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_temp3.jpg" width="375" style="margin:20"> </div> Result pair 4: <div align=center><img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_test4.jpg" width="375" style="margin:20"> &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; <img src="https://github.com/tangsanli5201/DeepPCB/blob/master/fig/result/result_temp4.jpg" width="375" style="margin:20"> </div> --> ##### Notification This work is contributed by the paper On-line PCB Defect Detector On A New PCB Defect Dataset. You can only use this dataset for research purposes. Contributor for paper with code: * Allena Venkata Sai Abhishek

Object Detection data set created from the engine DeepGTAV, which is based on the video game GTAV. Part of the three data sets proposed in the paper. This data set is motivated from the SeaDronesSee dataset with almost the same classes.

Object Detection data set created from the engine DeepGTAV, which is based on the video game GTAV. Part of the three data sets proposed in the paper. This data set is motivated from the VisDrone data set with almost the same classes.

A challenge that consists of three tasks, each targeting a different requirement for in-clinic use. The first task involves classifying images from the GI tract into 23 distinct classes. The second task focuses on efficiant classification measured by the amount of time spent processing each image. The last task relates to automatcially segmenting polyps. Source: EndoTect Challenge Please cite "The EndoTect 2020 Challenge: Evaluation andComparison of Classification, Segmentation and Inference Time for Endoscopy" if you use the dataset.

VD4UAV is an altitude-sensitive benchmark dataset designed to evade vehicle detection in Unmanned Aerial Vehicle (UAV) imagery. This dataset is specifically curated to facilitate the study of adversarial patch-based vehicle detection attacks in UAV images. The EVD4UAV dataset comprises a diverse set of images captured at various altitudes with fine-grained annotations, making it a robust platform for evaluating the performance of object detectors under adversarial conditions. Notably, the dataset includes around 3,000 images depicting winter scenarios where vehicles may be partially or fully covered by snow, providing a unique challenge for vehicle detection algorithms.

A 360-degree fisheye-like version of the popular FDDB face detection dataset. Source: FDDB-360: Face Detection in 360-degree Fisheye Images

This dataset is made up of forward-looking sonar images containing ten classes of underwater debris. The dataset can be used for segmentation or object detection. Applications include training computer vision models for underwater robotics applications.

FractureAtlas is a musculoskeletal bone fracture dataset with annotations for deep learning tasks like classification, localization, and segmentation. The dataset contains a total of 4,083 X-Ray images with annotation in COCO, VGG, YOLO, and Pascal VOC format. This dataset is made freely available for any purpose. The data provided within this work are free to copy, share or redistribute in any medium or format. The data might be adapted, remixed, transformed, and built upon. The dataset is licensed under a CC-BY 4.0 license. It should be noted that to use the dataset correctly, one needs to have knowledge of medical and radiology fields to understand the results and make conclusions based on the dataset. It's also important to consider the possibility of labeling errors. Furthermore, any publication that utilizes this resource should acknowledge the original paper, and the authors are encouraged to share their code and models to assist the research community in replicating the experiments and promoting the field of medical imaging.

FSVOD-500 is a large-scale video dataset comprising of 500 classes with class-balanced videos in each category for few-shot learning. FSVOD-500 is the first benchmark specially designed for few-shot video object detection for evaluating the performance of a given model on novel classes. Source: Few-Shot Video Object Detection Image source: Few-Shot Video Object Detection

The Human-Parts dataset is a dataset for human body, face and hand detection with ~15k images. It contains ~106k different annotations, with multiple annotations per image. Source: https://github.com/xiaojie1017/Human-Parts Image Source: https://github.com/xiaojie1017/Human-Parts

A Simulated Benchmark for multi-modal SLAM Systems Evaluation in Large-scale Dynamic Environments.

The INRIA-Horse dataset consists of 170 horse images and 170 images without horses. All horses in all images are annotated with a bounding-box. The main challenges it offers are clutter, intra-class shape variability, and scale changes. The horses are mostly unoccluded, taken from approximately the side viewpoint, and face the same direction. Source: Dynamical And-Or Graph Learning for Object Shape Modeling and Detection Image Source: http://calvin-vision.net/datasets/inria-horses/

Kvasir-Capsule dataset is the largest publicly released VCE dataset. In total, the dataset contains 47,238 labeled images and 117 videos, where it captures anatomical landmarks and pathological and normal findings. The results is more than 4,741,621 images and video frames altogether.

The LeukemiaAttri dataset is a large-scale, multi-domain collection of microscopy images derived from leukemia patient samples, enriched with detailed morphological information. This dataset comprises a total of 28.9K images (2.4K × 2 × 3 × 2), which were captured using both low-cost and high-cost microscopes at three different resolutions: 10x, 40x, and 100x, utilizing various cameras. In addition to providing location annotations for each white blood cell (WBC), the dataset includes comprehensive morphological attributes for every WBC, enhancing its utility for research and analysis in the field. | Sr. # | Dataset Name | Microscope | Resolution | Camera | Image Count | WBC Count | Artifacts Count | |-------|---------------|---------------|------------|----------------|-------------|-----------|-----------------| | 1 | H_100x_C1 | Olympus CX23 | 100x | Honor 9X Lite | 2.4k | 7.8k | 2.5k | | 2 | H_100x_C2 | Olympus CX23 | 100x | HD1500T | 2.4k | 7.8k | 2.5k | | 3 | H_40x_C1 | Olympus CX23 | 40x | Honor 9X Lite | 2.4k | 7.8k | 2.5k | | 4 | H_40x_C2 | Olympus CX23 | 40x | HD1500T | 2.4k | 7.8k | 2.5k | | 5 | H_10x_C1 | Olympus CX23 | 10x | Honor 9X Lite | 2.4k | 7.8k | 2.5k | | 6 | H_10x_C2 | Olympus CX23 | 10x | HD1500T | 2.4k | 7.8k | 2.5k | | 7 | L_100x_C1 | XSZ_107BN | 100x | Honor 9X Lite | 2.4k | 4.8k | 1.1k | | 8 | L_100x_C2 | XSZ_107BN | 100x | ZZCAT-5MP | 2.4k | 4.8k | 1.1k | | 9 | L_40x_C1 | XSZ_107BN | 40x | Honor 9X Lite | 2.4k | 7.8k | 2.5k | | 10 | L_40x_C2 | XSZ_107BN | 40x | ZZCAT-5MP | 2.4k | 7.8k | 2.5k | | 11 | L_10x_C1 | XSZ_107BN | 10x | Honor 9X Lite | 2.4k | 7.8k | 2.5k | | 12 | L_10x_C2 | XSZ_107BN | 10x | ZZCAT-5MP | 2.4k | 7.8k | 2.5k | | | Total Count | | | | 29k | 88.2k | 27.8k |

We introduce an object detection dataset in challenging adverse weather conditions covering 12000 samples in real-world driving scenes and 1500 samples in controlled weather conditions within a fog chamber. The dataset includes different weather conditions like fog, snow, and rain and was acquired by over 10,000 km of driving in northern Europe. The driven route with cities along the road is shown on the right. In total, 100k Objekts were labeled with accurate 2D and 3D bounding boxes. The main contributions of this dataset are: - We provide a proving ground for a broad range of algorithms covering signal enhancement, domain adaptation, object detection, or multi-modal sensor fusion, focusing on the learning of robust redundancies between sensors, especially if they fail asymmetrically in different weather conditions. - The dataset was created with the initial intention to showcase methods, which learn of robust redundancies between the sensor and enable a raw data sensor fusion in case of asymmetric sensor failure induced through adverse weather effects. - In our case we departed from proposal level fusion and applied an adaptive fusion driven by measurement entropy enabling the detection also in case of unknown adverse weather effects. This method outperforms other reference fusion methods, which even drop in below single image methods. - Please check out our paper for more information.

This data set contains 775 video sequences, captured in the wildlife park Lindenthal (Cologne, Germany) as part of the AMMOD project, using an Intel RealSense D435 stereo camera. In addition to color and infrared images, the D435 is able to infer the distance (or “depth”) to objects in the scene using stereo vision. Observed animals include various birds (at daytime) and mammals such as deer, goats, sheep, donkeys, and foxes (primarily at nighttime). A subset of 412 images is annotated with a total of 1038 individual animal annotations, including instance masks, bounding boxes, class labels, and corresponding track IDs to identify the same individual over the entire video. Source: Lindenthal Camera Traps on lila.science Image Source: Lindenthal Camera Traps on lila.science

• 5 domains: synthetic domain, document domain, street view domain, handwritten domain, and car license domain * over five million images

The MultiviewC dataset mainly contributes to multiview cattle action recognition, 3D objection detection and tracking. We build a novel synthetic dataset MultiviewC through UE4 based on real cattle video dataset which is offered by CISRO. The format of our data set has been adjusted on the basis of MultiviewX for set-up, annotation and files structure.

Occluded COCO is automatically generated subset of COCO val dataset, collecting partially occluded objects for a large variety of categories in real images in a scalable manner, where target object is partially occluded but the segmentation mask is connected.