Hysen Labscascade-detr
[ICCV'23] Cascade-DETR: Delving into High-Quality Universal Object Detection
Cascade DETR: object detection from ICCV 2023 with UDB10
An ICCV 2023 detection method that targets cross domain generalization and localization accuracy at the same time, shipped with code, ten public datasets, and pretrained checkpoints.
What it proposes
The paper, published at ICCV 2023, tackles two problems in one architecture. Generalization across diverse domains and localization accuracy are handled together through a Cascade Attention layer that limits attention to the previous box prediction, pulling object centric information into the detection decoder. Query scoring also changes: instead of relying on classification scores, the model predicts the expected IoU of a query, which the authors say produces better calibrated confidences.
The UDB10 benchmark
Alongside the method comes a benchmark called UDB10, built from 10 datasets across diverse domains. The authors report that Cascade DETR advances the state of the art on COCO and improves DETR based detectors on every UDB10 dataset, in some cases by more than 10 mAP. Gains under strict quality requirements are described as even more pronounced.
Installing the dependencies
Setup is standard for this kind of research code. Both PyTorch and TorchVision need to be installed with CUDA support. That is the whole environment requirement stated in the README.
Getting the datasets
All ten datasets download from public links, and the paper reports separate training results for each one. Some datasets come from Roboflow, where the README recommends the COCO JSON download format, while others come from official sites, such as Cityscapes and BDD100K, with pre converted COCO jsons offered for the latter to skip the transformation step. Expected folder structures are described so the data matches the training scripts.
Training and evaluation
The getting started section points to a project folder with detailed training and evaluation scripts and pretrained checkpoints, so results can be reproduced without training everything from scratch. The README closes by asking anyone who finds Cascade DETR useful in their research, or who refers to the baseline results, to star the repository and consider citing the paper.
Community notes