Fetch and Forge: Efficient Dataset Condensation for Object Detection

Authors: Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, Cairong Zhao

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various detection datasets demonstrate the superiority of DCOD. Even at an extremely low compression rate of 1%, we achieve 46.4% and 24.7% AP50 on the VOC and COCO, respectively, significantly reducing detector training duration.
Researcher Affiliation Collaboration Ding Qi1 Jian Li2 Jinlong Peng2 Bo Zhao4 Shuguang Dou1 Jialin Li2 Jiangning Zhang2 Yabiao Wang3,2 Chengjie Wang2 Cairong Zhao1 1Tongji University 2Tencent You Tu Lab 3Zhejiang University 4Shanghai Jiao Tong University
Pseudocode Yes Algorithm 1 Dataset Condensation for Object Detection
Open Source Code No 5. Open access to data and code...Answer: [No] Justification: We commit to releasing the complete code later.
Open Datasets Yes Our DCOD method is evaluated on Pascal VOC [7, 6] and MS COCO [13] benchmarks, with image resolution set to 512x512.
Dataset Splits No The paper describes train/test splits for Pascal VOC and MS COCO, for example, "For Pascal VOC, we merge the trainval sets of VOC2007 and VOC2012 into a single training set with 16,551 images, using the VOC2007 test set for evaluation." While it mentions "validation set" in the context of meta-learning frameworks (Figure 1a), it does not specify a validation split or how it was used in its own experimental setup for DCOD.
Hardware Specification Yes All experiments were performed on a single V100 gpu.
Software Dependencies No The paper mentions models like YOLOv3-SPP and Faster R-CNN, and a feature extractor ResNet50. However, it does not provide specific version numbers for these software components or any other libraries (e.g., PyTorch, TensorFlow, Python version) used in the experiments.
Experiment Setup Yes During the standard image synthesis process, we set the image learning rate at 0.002. The weight for the task loss is set at 1, while Rfeatureis assigned a weight of 0.1. The αT V and αl2 is established at 1 and 0.001, respectively. The α in background suppression strategy is set to 0.7.