Final integration projects from Images and Vision: From Pixels to Deep Learning (BIU, 2026). Each team combines the whole course into one study: benchmark classical and deep vision methods, degrade the input with realistic distortions, and measure how much image restoration and model fine-tuning recover. See the project requirements and assessment criteria.
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Evaluates the robustness of classical low/high-level vision and a deep detector for autonomous driving as image quality degrades across nine SNR-quantified severity levels, then tests two recovery strategies: classical pre-processing enhancement and model fine-tuning.
Amit Ashkenazi, Alon Kenan
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Runs detection, drivable-area segmentation, and scene classification on the same BDD100K images and measures how each degrades under noise, blur, and rain. Each distortion is paired with a matched classical enhancement, and the deep models are also fine-tuned to test recovery.
Matan Lerner, Ofek Nasi
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Benchmarks three deep models at increasing levels of scene understanding (boxes, instance masks, dense depth) as images degrade, then compares classical image enhancement against fine-tuning as recovery strategies, reporting RMSE/SSIM, IoU, and recall.
Noa Speyer, Ayelet Levy
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Measures how two classical vision tasks and a YOLOv8n detector degrade under JPEG, low-light, and motion blur on real BDD100K images, then compares classical restoration (including three motion-blur deconvolution methods) against detector fine-tuning for recovery.
Elad Havakuk
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Compares one classical method (ORB matching) against two pretrained deep models on PASCAL VOC as images are progressively degraded, adding an enhancement stage, fine-tuning, and a gradual SNR sweep to probe model limits.
Eve Yatzkan, Maayan Zelig
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A four-task robustness study on 30 COCO images evaluating classical (ORB, Canny) and deep (YOLOv8n, SegFormer-B0) methods, pairing each distortion with a matched classical enhancement and quantifying degradation and recovery per task.
Gilad Korengut
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Benchmarks low- to high-level vision tasks on the Oxford-IIIT Pet dataset under salt-and-pepper, ISO noise, and fog, comparing classical pre-processing (adaptive median/bilateral filters, dark-channel-prior dehazing) against fine-tuning the deep models on degraded data.
Ido Ben David, Shoval Bracha
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A cross-domain study running the full clean/distort/enhance/fine-tune chain on ADE20K (indoor) and KITTI (driving) so findings validate across domains, recovering performance through classical restoration and deep model fine-tuning.
Malak Mrowat, Jobran Khateeb
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
An SNR-parameterized robustness study across COCO-2017 and ADE20K, comparing classical enhancement filters and a learned ResNet-18 restoration model against deep-model fine-tuning as recovery strategies.
Eran Guetta, Omer Lazaros
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A medical-imaging robustness study on CVC-ClinicDB colonoscopy polyp frames, testing ORB, YOLOv8n, and SegFormer-B0 under noise, low contrast, and JPEG against PSNR, with recovery via classical enhancement (NLM, CLAHE, bilateral) and YOLO fine-tuning.
Lior Avrahami
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A three-level human-perception robustness study (Canny boundary, YOLOv8n detection, pose estimation) on a COCO-person subset under surveillance-style degradations, measuring recovery via classical enhancement versus corruption-aware pose fine-tuning.
Michal Laufer
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
An autonomous-driving robustness study asking whether degradation should be fixed in the image (classical restoration) or in the model (fine-tuning), evaluated across three vision tasks at four intensity levels; fine-tuning recovers most of the lost performance.
Alon Ron, Nadav Moshe
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Evaluates the robustness of classical and deep computer-vision approaches under common distortions and investigates whether image enhancement can recover lost performance, via a modular notebook pipeline over a COCO subset.
Vered Sabban, Nithay Cohen
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Evaluates low-level and high-level vision algorithms under three distortions and studies two mitigation approaches (pre-processing enhancement and fine-tuning); a bilateral filter recovers YOLO recall from 0.067 to 0.633 under JPEG compression.
Niv Orkabi, Dareen Sawaed
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A project spanning four vision tasks and four distortions at four SNR-quantified severity levels, distinguishing activity metrics from correctness and comparing classical enhancement against detector fine-tuning for recovery.
Nitzan Sharabi, Roni Volshtein, Matan Sella
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Systematically benchmarks four vision tasks on a large fixed COCO subset, comparing classical restoration (BM3D, median, Wiener) against deep fine-tuning using the official COCO and Panoptic APIs for reproducible AP, OKS, and PQ metrics.
Shira Tziony, Ohad Shpizhizen
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Evaluates classical algorithms and a deep model across low-, mid-, and high-level vision tasks under three distortions, measuring clean baselines, degradation, classical restoration, and distortion-aware fine-tuning of ResNet-50.
Hadar Cemama, Tehila Segal
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Tests whether fixing the image (classical restoration) or fixing the model (fine-tuning) better recovers vision performance under three distortions on a frozen DOTA-v1.0 subset; haze favors dehazing, noise favors fine-tuning, JPEG is a tie.
Yair Nachum, Yacov Yoles
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
Measures how noise, blur, and JPEG affect low-level feature detection (SIFT) and high-level scene understanding (ResNet-50, DeepLabV3) on Oxford-IIIT Pet; whether cleaning the image or adapting the model wins depends on the task's level of abstraction.
Yonatan Haba, Elad Tsachi
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A multi-tier robustness audit of ORB, K-Means, and YOLOv8 on KITTI driving imagery with dense SNR sweeps, contrasting classical restoration preprocessing against seven YOLOv8 fine-tuning strategies (plus a curriculum schedule) and a hybrid of both.
Ziv Chaba
Images and Vision: From Pixels to Deep Learning · BIU EE 2026
A study of how ORB, Canny, SegFormer-B0, and YOLOv8n behave on Cityscapes under four distortions, with paired-bootstrap statistics, classical restoration, and distortion-aware YOLO fine-tuning, keeping negative gains visible rather than filtered.
Nir Altahan, Ofek Valdman, Arik Slavsky
