Sparse Semi-DETR is a novel transformer-based, end-to-end semi-supervised object detection method that addresses the limitations of DETR-based semi-supervised object detection (SSOD). The method introduces two key modules: a Query Refinement Module and a Reliable Pseudo-Label Filtering Module. The Query Refinement Module enhances the quality of object queries by fusing low-level and high-level features from weakly augmented images, improving detection of small and partially obscured objects. The Reliable Pseudo-Label Filtering Module selectively filters high-quality pseudo-labels, enhancing detection accuracy and consistency. Sparse Semi-DETR achieves significant improvements on the MS-COCO and Pascal VOC benchmarks, outperforming state-of-the-art methods, particularly in detecting small or partially obscured objects. The method is evaluated on three scenarios: COCO-Partial, COCO-Full, and VOC, demonstrating superior performance across all settings. Sparse Semi-DETR outperforms existing SSOD methods in terms of detection accuracy and efficiency, with results showing a 44.3 mAP on MS-COCO using only 10% labeled data. The method also shows improved performance on the Pascal VOC benchmark, surpassing supervised baselines by significant margins. The paper presents ablation studies showing that both modules contribute to the method's effectiveness, with the Query Refinement Module significantly improving performance. The method is designed to be compatible with various DETR-based detectors, offering flexibility in integration. Overall, Sparse Semi-DETR provides a more efficient and accurate approach to semi-supervised object detection, particularly in challenging scenarios involving small or partially obscured objects.Sparse Semi-DETR is a novel transformer-based, end-to-end semi-supervised object detection method that addresses the limitations of DETR-based semi-supervised object detection (SSOD). The method introduces two key modules: a Query Refinement Module and a Reliable Pseudo-Label Filtering Module. The Query Refinement Module enhances the quality of object queries by fusing low-level and high-level features from weakly augmented images, improving detection of small and partially obscured objects. The Reliable Pseudo-Label Filtering Module selectively filters high-quality pseudo-labels, enhancing detection accuracy and consistency. Sparse Semi-DETR achieves significant improvements on the MS-COCO and Pascal VOC benchmarks, outperforming state-of-the-art methods, particularly in detecting small or partially obscured objects. The method is evaluated on three scenarios: COCO-Partial, COCO-Full, and VOC, demonstrating superior performance across all settings. Sparse Semi-DETR outperforms existing SSOD methods in terms of detection accuracy and efficiency, with results showing a 44.3 mAP on MS-COCO using only 10% labeled data. The method also shows improved performance on the Pascal VOC benchmark, surpassing supervised baselines by significant margins. The paper presents ablation studies showing that both modules contribute to the method's effectiveness, with the Query Refinement Module significantly improving performance. The method is designed to be compatible with various DETR-based detectors, offering flexibility in integration. Overall, Sparse Semi-DETR provides a more efficient and accurate approach to semi-supervised object detection, particularly in challenging scenarios involving small or partially obscured objects.