AE Pixel Sorter was the first tool to bring the pixel sorting glitch effect available for motion designers and video editors within After Effects. Experimenting with Pixel Sorting just became so much easier! Since its launch, AEPS had a huge success in Broadcast TV, Live performances, VJs, Music Videos, Video Games, Photography and even Apparel!
AE Pixel Sorter was the first tool to bring the pixel sorting glitch effect available for motion designers and video editors within After Effects. Experimenting with Pixel Sorting just became so much easier.
It is worth noting that some existing encryption systems are still vulnerable to cracking due to insufficient consideration of the chaotic characteristics of the system and the lack of security of the algorithm itself in the scheme design . Dhall et al. made a cryptographic analysis of the image encryption scheme proposed in literature  and found some problems and unenforceability in the encryption scheme. Finally, they improved the scheme effectively to improve the security of the algorithm. In order to avoid such a situation, this paper adopts the dynamic random diffusion method based on Hilbert curve in the diffusion process when designing the encryption algorithm. When changing the size of the pixel value, the pixel position also changes, which not only improves the efficiency of the algorithm, but also greatly improves the security of the algorithm.
Step 2. The pixel matrix represented by the original image IM is mapped in a two-dimensional rectangular coordinate system. As shown in Figure 11, (50,97,112) marked in red indicates that the pixel value of the image matrix at (50,97) is 112 in a two-dimensional rectangular coordinate system.
Step 2. The sum of pixel values SUM1,SUM2,SUM3 and SUM4 of image IM1,IM2,IM3,IM4, as well as their information entropy KS,KS1,KS2,KS3,KS4 are calculated respectively, as shown in Equations (9) and (10):
Step 3. The key stream generation process embeds parameters related to plaintext, which can effectively resist common known/chosen plaintext attacks. We use plaintext related pixel information SUM1, SUM2, SUM3, SUM4, and information entropy KS,KS1,KS2,KS3,KS4 to generate the initial value of the chaotic system. The method is given in Equation (11). In this way, when inputting different original images for encryption, the system will generate completely different random key streams. This greatly enhances the security of the algorithm:
Step 3. The index matrix ID1 is used to perform further column scrambling on the pixel matrix SCR after fractal-like model scrambling. The pixel matrix ISC1 after column scrambling is generated by Equation (16):
where n is pair numbers of adjacent pixels, pi and qi are a pair of adjacent pixel values, E(p) is the mean of p, E(q) is the mean of q, D(p) is the variance of p, D(q) is the variance of q, and cov(p,q) represents the covariance of p and q. Randomly select 5000 pairs of adjacent pixels in each direction from the image samples used in this algorithm and their encrypted images. The correlation coefficients are calculated and compared with other advanced references, and the results are listed in Table 3. As can be seen from Table 3, the correlation between adjacent pixels in each direction in the plain images is strong, and the correlation coefficients are all close to 1. However, in the proposed encryption algorithm, the correlation coefficient of ciphertext image is closer to 0, which indicates that the correlation between adjacent pixels of ciphertext image can be ignored.
where xi represents the i-th pixel value, and p(xi) represents the probability of pixel xi. Table 4 is the calculation results and comparative analysis of the information entropy of the image sample in this experiment. The information entropy of ciphertext image is close to 8. Compared with other algorithms, the proposed algorithm has greater information entropy. This shows that the encrypted image is extremely random, which is difficult to crack.
Object detection is an important research topic in the computer vision field with many real-life applications such as athlete detection for sport performance analysis , pedestrian detection for automated vehicles , intruder detection for security surveillance systems , concrete crack detection , and geospatial image analysis (a particularly challenging application of object detection to optical remote sensing ). With recent advancements in the field of deep learning, detection performance has improved to a great extent. However, object detection is still a challenging problem in crowded scenes where multiple objects tend to occlude each other. In such scenarios established detectors often fail to detect the overlapping objects.
Our proposed bounding box selection algorithm selects bounding boxes based on a predicted overlap map generated from ground truth bounding box annotations. An overlap map is a 2D heatmap equal in size to the original image where each pixel of the overlap map determines the number of bounding boxes intersecting the pixel. Figure 1b shows an image and corresponding overlap map. For example, it shows the pixel value is three in the shared region of the three overlapping people. Since ground truth bounding boxes are not available at inference time, we create an overlap map prediction model to predict overlap maps. In particular, we use the semantic segmentation model, deeplabv3+  as the backbone of our proposed overlap map model.
To evaluate the performance of our overlap-aware bounding box selection we conducted experiments on two datasets (the crowdHuman dataset  and a sports dataset) which contain frequently overlapping objects. Using a state-of-the-art prediction model , we produced candidate object detections. Separately, we trained a model to produce overlap maps from images. We then applied the pixel voting bounding box selection algorithm to select actual bounding boxes from duplicate predictions using overlap map guidance.
Our results on both person-detection datasets demonstrate that our proposed solution achieves better results than the existing set-NMS approach in terms of localisation-precision-recall (LRP)  detection metrics. We also show that our pixel voting bounding box selection accuracy improves drastically when ground truth overlap maps are used which shows the potential of our method to offer even better performance if a better overlap map prediction is developed.
We perform an extensive experimental study involving two different datasets to compare the performance of our pixel voting bounding box selection algorithm with the existing state-of-the-art set-NMS  based approach. The results show overlap-aware pixel voting outperforms the set-NMS approach.
Han et al.  propose an end-to-end object relation model that prunes duplicate bounding boxes without the need for any further post-processing. It uses an attention module that captures the relationships between objects based on their appearance and location. The relation module is used for predicting the initial set of bounding boxes and also used for duplicate removal. Recently Zixuan et al.  proposed a method that models each pedestrian using a beta distribution within the bounding box of the pedestrian. This allows higher probability to be assigned to the pixels of the pedestrian within the bounding box rather than assigning equal probability to the pixels in the bounding box. Their method performs bounding box selection using a method called BetaNMS which uses KL divergence between the beta distribution of pairs of bounding boxes to determine how likely two bounding boxes correspond to the same person. Both  and  box selection algorithms are tightly integrated with the detector itself making them incompatible with current and future alternative object detectors. In contrast, our box selection algorithm is decoupled from the object detector.
The key component of our bounding box selection algorithm is the overlap map. The overlap map represents the number of objects overlapping each pixel of the image. Predicting an overlap map is a very similar problem to semantic segmentation where the task is to label each pixel in terms of which class it belongs to. Therefore most existing semantic segmentation solutions can be used to output the overlap map.
Instance segmentation [9, 13] and panoptic segmentation [2, 21] problems are also related to our overlap map prediction problem. Instance segmentation is the task of detecting each distinct object of interest appearing in an image. Lei and Tai et al.  proposed a deep occlusion-aware instance segmentation model that uses different layers for occluding and occluded objects. This model requires annotation of both occluding and occluded objects explicitly. Recent research introduced panoptic segmentation, a new segmentation task by combining both semantic and instance segmentation. Existing work [2, 21] proposed panoptic segmentation that does per pixel instance classification without outputting the bounding box of the objects. In this task, objects are only represented as irregular shapes that are defined by the visible pixels. 153554b96e