Muhammad Naeem

Muhammad Naeem
Graduate Student in LIESMARS
Wuhan University, Hubei, China
Advisor: Prof. Xiongwu Xiao

Email: m.naeem4288@gmail.com
LinkedIn: linkedin.com/in/muhammadnaeem27
Portfolio: muhammadnaeem27.github.io

• Implemented an inpainting approach using Stable Diffusion with ControlNet to furnish empty room images without altering existing items and color combinations.
• Applied semantic segmentation on empty rooms to identify areas and items to preserve, generating binary images from segmented masks.
• Used tailored prompts for bedrooms and living rooms, passing binary masked images, prompts, and empty room images to the Stable Diffusion inpainting model to generate furniture in the specified areas.

• Developed a system to generate interesting highlights from 6-7 hours of video game streaming videos using three main approaches: Sound, Transcription, and Movement.
• Implemented separate AI techniques for each approach to identify key timestamps, leveraging audio analysis, natural language processing, and computer vision.
• Combined results from all three approaches in a specific sequence to produce the best highlights and generate a final video.

• Compiled a diverse dataset comprising four categories of weed. Implemented YOLO-v5 for real-time object detection, achieving precise identification of weeds in agricultural fields.
• Designed and built a robotic car prototype that uses a smartphone camera and DC motors to wirelessly stream live images to the YOLO-v5 model via FastAPI.
• Integrated an Arduino-controlled LED system with a matrix-based system to precisely locate and eliminate detected weeds.
• Created a website to showcase project insights, allowing users to engage with the trained YOLO-v5 model and evaluate its effectiveness.
• Used data visualization to highlight the project's technological innovation and potential for sustainable agricultural practices.

• Used U-Net structures for both rough and detailed portrait segmentation, addressing the problem of subject isolation in images.
• Utilized internet and publicly available datasets like COCO for training data.

• Included two time-scale update rules, self-attention GANs, and inflection points for effective and error-free coloring.
• Used U-Net architecture with model-specific backbone choices (resnet34/resnet101), spectral normalization, and self-attention.
• Implemented Perceptual Loss based on VGG16 during NoGAN learning for realistic colorization.

• Created talking videos by combining lips and head movements of a reference video with a source image using Dlib for facial landmark identification.
• Used Wav2Lip-hq for precise lip movement synchronization and ESRGAN for video up-sampling.
• Integrated synchronized audio and used ffmpeg and OpenCV for efficient video processing.

• Created a Document Scanner with a Geometric Unwarping Transformer using Doc3D dataset for accurate geometric unwrapping.
• Used DocProj dataset for Illumination Correction Transformer to resolve illumination problems during document scanning.

• Developed a model with sky-changing (Multiband blender), high-res processing (UNET), and low-res processing (hrnet-ocr).
• Enhanced segmentation in the low-res module using ASPP and improved high-res module with learnable parameters.
• Used Multiband blender for effective sky replacement.

• Implemented Stable Diffusion for text-to-image generation, using fine-tuning and Dreambooth for high-quality, customized images.

• Implemented pose-matching and facial point detection for accurate face swapping using Dlib.
• Blended faces to ensure natural skin color transitions.