Building Blocks for Robust and Effective Semi-Supervised Real-World Object Detection

Authors: Moussa Kassem Sbeyti, Nadja Klein, Azarm Nowzad, Fikret Sivrikaya, Sahin Albayrak

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our methods through comprehensive experiments on autonomous driving datasets, resulting in up to 6% increase in SSOD performance. Overall, our investigation and novel, data-centric, and broadly applicable building blocks enable robust and effective SSOD in complex, real-world scenarios. Code is available at https://mos-ks.github.io/publications.
Researcher Affiliation Collaboration Moussa Kassem Sbeyti1,2,3 Nadja Klein1 Azarm Nowzad3 Fikret Sivrikaya2 Sahin Albayrak2 1Scientific Computing Center, Karlsruhe Institute of Technology 2DAI-Labor, Technische Universität Berlin 3Continental AG EMAIL EMAIL EMAIL
Pseudocode No The paper describes its methods (RCC, RCF, GLC, PLS) using descriptive text and mathematical formulas within the main body, but does not present any distinct pseudocode blocks or algorithms.
Open Source Code Yes Code is available at https://mos-ks.github.io/publications.
Open Datasets Yes Experiments are conducted on two autonomous driving datasets: KITTI (Geiger et al., 2012) (7 classes, 20% random split for validation) and BDD100K (Yu et al., 2020) (10 classes, 12.5% official split).
Dataset Splits Yes Experiments are conducted on two autonomous driving datasets: KITTI (Geiger et al., 2012) (7 classes, 20% random split for validation) and BDD100K (Yu et al., 2020) (10 classes, 12.5% official split).
Hardware Specification No The paper does not specify the exact hardware (e.g., GPU models, CPU types) used for running the experiments. It only mentions 'efficient GPU utilization' in the context of batch size.
Software Dependencies No To demonstrate the effectiveness of our data-centric methods, we select Efficient Det-D0 (Tan et al., 2020; Google, 2020) pre-trained on MS COCO (Lin et al., 2014) as our baseline detector. ... As an exemplary student-teacher framework, this work adopts STAC (Sohn et al., 2020) ... For data augmentations (A), we use Rand Augment (Cubuk et al., 2020). ... we employ the Tensor Flow implementation of YOLOv3 (Farhadi & Redmon, 2018; Zhang, 2019). No specific version numbers for these software components or their underlying frameworks (e.g., TensorFlow, PyTorch) are provided.
Experiment Setup Yes Each training comprises 200 epochs with 8 batches and an input image resolution of 1024 512 pixels. All other hyperparameters of Efficient Det are set to their default values (Tan et al., 2020). ... As an exemplary student-teacher framework, this work adopts STAC (Sohn et al., 2020) with λ = 1 in Eq. (1)... We fine-tune a YOLOv3 model pretrained on MS COCO using 10% of KITTI for 50 epochs with an input resolution of 1024 1024, a batch size of 8, a learning rate of 0.0001, and anchors optimized via k-means clustering. All other hyperparameters remain at their default values.