cascade_rcnn

Cascade R-CNN

Cascade R-CNN: High Quality Object Detection and Instance Segmentation

Abstract

In object detection, the intersection over union (IoU) threshold is frequently used to define positives/negatives. The threshold used to train a detector defines its quality. While the commonly used threshold of 0.5 leads to noisy (low-quality) detections, detection performance frequently degrades for larger thresholds. This paradox of high-quality detection has two causes: 1) overfitting, due to vanishing positive samples for large thresholds, and 2) inference-time quality mismatch between detector and test hypotheses. A multi-stage object detection architecture, the Cascade R-CNN, composed of a sequence of detectors trained with increasing IoU thresholds, is proposed to address these problems. The detectors are trained sequentially, using the output of a detector as training set for the next. This resampling progressively improves hypotheses quality, guaranteeing a positive training set of equivalent size for all detectors and minimizing overfitting. The same cascade is applied at inference, to eliminate quality mismatches between hypotheses and detectors. An implementation of the Cascade R-CNN without bells or whistles achieves state-of-the-art performance on the COCO dataset, and significantly improves high-quality detection on generic and specific object detection datasets, including VOC, KITTI, CityPerson, and WiderFace. Finally, the Cascade R-CNN is generalized to instance segmentation, with nontrivial improvements over the Mask R-CNN.

Results and Models

Cascade R-CNN

Backbone	Style	Lr schd	Mem (GB)	Inf time (fps)	box AP	Config	Download
R-50-FPN	caffe	1x	4.2		40.4	config	model | log
R-50-FPN	pytorch	1x	4.4	16.1	40.3	config	model | log
R-50-FPN	pytorch	20e	-	-	41.0	config	model | log
R-101-FPN	caffe	1x	6.2		42.3	config	model | log
R-101-FPN	pytorch	1x	6.4	13.5	42.0	config	model | log
R-101-FPN	pytorch	20e	-	-	42.5	config	model | log
X-101-32x4d-FPN	pytorch	1x	7.6	10.9	43.7	config	model | log
X-101-32x4d-FPN	pytorch	20e	7.6		43.7	config	model | log
X-101-64x4d-FPN	pytorch	1x	10.7		44.7	config	model | log
X-101-64x4d-FPN	pytorch	20e	10.7		44.5	config	model | log

Cascade Mask R-CNN

Backbone	Style	Lr schd	Mem (GB)	Inf time (fps)	box AP	mask AP	Config	Download
R-50-FPN	caffe	1x	5.9		41.2	36.0	config	model | log
R-50-FPN	pytorch	1x	6.0	11.2	41.2	35.9	config	model | log
R-50-FPN	pytorch	20e	-	-	41.9	36.5	config	model | log
R-101-FPN	caffe	1x	7.8		43.2	37.6	config	model | log
R-101-FPN	pytorch	1x	7.9	9.8	42.9	37.3	config	model | log
R-101-FPN	pytorch	20e	-	-	43.4	37.8	config	model | log
X-101-32x4d-FPN	pytorch	1x	9.2	8.6	44.3	38.3	config	model | log
X-101-32x4d-FPN	pytorch	20e	9.2	-	45.0	39.0	config	model | log
X-101-64x4d-FPN	pytorch	1x	12.2	6.7	45.3	39.2	config	model | log
X-101-64x4d-FPN	pytorch	20e	12.2		45.6	39.5	config	model | log

Notes:

• The 20e schedule in Cascade (Mask) R-CNN indicates decreasing the lr at 16 and 19 epochs, with a total of 20 epochs.

Pre-trained Models

We also train some models with longer schedules and multi-scale training for Cascade Mask R-CNN. The users could finetune them for downstream tasks.

Backbone	Style	Lr schd	Mem (GB)	Inf time (fps)	box AP	mask AP	Config	Download
R-50-FPN	caffe	3x	5.7		44.0	38.1	config	model | log
R-50-FPN	pytorch	3x	5.9		44.3	38.5	config	model | log
R-101-FPN	caffe	3x	7.7		45.4	39.5	config	model | log
R-101-FPN	pytorch	3x	7.8		45.5	39.6	config	model | log
X-101-32x4d-FPN	pytorch	3x	9.0		46.3	40.1	config	model | log
X-101-32x8d-FPN	pytorch	3x	12.1		46.1	39.9	config	model | log
X-101-64x4d-FPN	pytorch	3x	12.0		46.6	40.3	config	model | log

Citation

@article{Cai_2019,
   title={Cascade R-CNN: High Quality Object Detection and Instance Segmentation},
   ISSN={1939-3539},
   url={http://dx.doi.org/10.1109/tpami.2019.2956516},
   DOI={10.1109/tpami.2019.2956516},
   journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
   publisher={Institute of Electrical and Electronics Engineers (IEEE)},
   author={Cai, Zhaowei and Vasconcelos, Nuno},
   year={2019},
   pages={1–1}
}