Backgroound Image

Month: May 2025

Editing safeguards update

Posted on by Roshinka

In 2022 and 2023, as a shield against unregulated LLM’s running amok in this world, I was using an enormous amount of animated pattern overlays, filters, and stickers on top of original footage in addition to significantly downgrading footage quality to blur biometric features and feed AI mainly erroneous and useless information. 

In 2024, when the majority of users footage and data has already been unlawfully assimilated by the big AI companies, I’ve implemented other methods for disrupting LLM’s “learning” that I will discuss in 2026. 

In 2025, I aim for even more seamless LLM spoofing. As already mentioned, silencing my online activity forever is on the table.

And maybe it goes on and on, and on and on…

Posted on by Roshinka

Scream and Shout, Just Dance Extreme, 15s preview. Shot at 4K, 24fps and intentionally downgraded to 1080p.

I started working on a new editing software with strong AI features, but I’m strictly cutting for now, not using effects yet. I’ve wanted to use this app for a long time, but couldn’t for legal reasons. Now that they are subject to at least one class action and likely a ban, things are bit more nuanced, not necessarily safe but tolerable.

UK slays at Eurovision with What The Hell Just Happened

Posted on by Roshinka

this song is really well written (like listening to Queen but more current) and the performance is out of this world. This is definitely my most favorite song from the entire Eurovision. I am going to listen to it on repeat for a little while. I was also impressed by Sweden (Bara Bada Bastu, total vibe) and Switzerland (Voyage, cinematic emotional). Now I have a playlist of 6. 

 

Azerbaijan at Eurovision, Run With You

Posted on by Roshinka

that gadulka shashtar solo takes the cake. I had to research the exact instrument because they have similar shaped ones with different number of strings throughout Eastern Europe and the Eurasian republics. I think this is the tar or shashtar, but I’ll obsess over it until I’m certain. So far I have three favorite songs from Eurovision, this one from Azerbaijan; Iceland, VÆB – RÓA which also has an epic trad violin solo; and Australia, Milkshake Man for the style and humour.

K-Pop Dance Pose Detection AI Model

Posted on by Roshinka

Speaking of ideas that want to be developed, I’m still waiting for a viable pose detection model to be made. It would be nice to take any dance practice video and ask AI to break down the moves per dancer and make a tutorial, like a dance prof would do. Not like I haven’t asked. At this point any rudimentary model can give you accurate legal information from any jurisdiction and any period of time. Constitution, piece of cake. Criminal law, litigation, novel legal questions, all covered. When it comes to dance however, seems like the hardest thing to do.

Since the paper on Comparative K-Pop Choreography Analysis Through Deep Learning Pose Estimation across a Large Video Corpus came out in 2021 in the Digital Humanities Quarterly, I haven’t seen any advancement. 

Since then, the video corpus of dance practice has considerably expanded thanks to humans and it should be so much easier for AI, right? Right?!?

Crickets.

the number of humans you will need to annotate frames to make datasets for the model to run would exceed the number of humans on this planet. 

There are AI-based apps from 3 years ago like SyncTrainer that allow you to upload your performance to compare of a K-pop chorus with the original dance practice to tell you approximately how well you did, but this is still on the level where humans get to do all the work. To use the app, I need to first detect all the poses myself, learn them, perform them, film them, upload them, to compare the final result. Quite a nuance. I want an app that can be a dance instructor.

here is where pose estimation is in 2025, (https://viso.ai/deep-learning/pose-estimation-ultimate-overview/ ), so pretty much nowhere. Anyone who tells you that AI can replace humans is misinformed. I personally ain’t gonna lift a finger to change that. 

AI’s defense regarding the main challenges of pose detection

Detecting the human pose is a challenging task because the body’s appearance changes dynamically due to diverse forms of clothing, arbitrary occlusion, occlusions due to the viewing angle, and background contexts. Pose estimation needs to be robust to challenging real-world variations such as lighting and weather.

Therefore, it is challenging for image processing models to identify fine-grained joint coordinates. It is especially difficult to track small and barely visible joints.

The main challenges of pose detection in AI and computer vision include:

  • Occlusions: When body parts are hidden or blocked by other objects or people, it becomes difficult for the model to accurately detect and track keypoints. This is a significant issue in crowded scenes or when people are interacting closely.

  • Crowded Scenes: Detecting and tracking multiple individuals in crowded environments is complex. The model must distinguish between different people and their respective keypoints, which can be challenging when individuals overlap or are in close proximity.

  • Motion Blur and Defocus: In video sequences, motion blur and defocused video frames can degrade the performance of pose estimation models. These issues are particularly problematic for real-time tracking, where the model needs to handle dynamic and fast-moving scenarios.

  • Temporal Dependency: Capturing the temporal dependency among video frames is crucial for accurate multi-frame human pose estimation. However, this is a challenging task, especially when dealing with high-speed movements or complex interactions.

  • Variations in Body Size and Pose: Pose estimation models must account for variations in body size and pose. For example, some bodies may be farther from the camera, making certain body parts almost invisible and difficult to estimate.

  • Lighting Conditions and Backgrounds: Different environments with varying light contrast and complex backgrounds can affect the accuracy of pose estimation. The model needs to be robust enough to handle these variations.

  • Limited Data and Annotated Datasets: Training pose estimation models requires large and diverse annotated datasets. For highly specific movements, such as those in rehabilitation or sports, custom data is often necessary, which can be time-consuming and costly to obtain.

  • Model Complexity and Performance: More advanced pose estimation models with high accuracy can be computationally expensive and have a large memory footprint. Balancing model complexity and performance is a challenge, especially for deployment on various devices with different computational capabilities.

  • Privacy Concerns: While privacy concerns are generally less relevant in human pose estimation compared to other computer vision tasks, they can still arise, particularly in applications involving surveillance or personal data.

Addressing these challenges requires ongoing research and development in areas such as dataset creation, model architecture, and algorithm optimization.