Recursively Refined R-CNN: Instance Segmentation with Self-RoI Rebalancing

3 Apr 2021  ·  Leonardo Rossi, Akbar Karimi, Andrea Prati ·

Within the field of instance segmentation, most of the state-of-the-art deep learning networks rely nowadays on cascade architectures, where multiple object detectors are trained sequentially, re-sampling the ground truth at each step. This offers a solution to the problem of exponentially vanishing positive samples. However, it also translates into an increase in network complexity in terms of the number of parameters. To address this issue, we propose Recursively Refined R-CNN (R^3-CNN) which avoids duplicates by introducing a loop mechanism instead. At the same time, it achieves a quality boost using a recursive re-sampling technique, where a specific IoU quality is utilized in each recursion to eventually equally cover the positive spectrum. Our experiments highlight the specific encoding of the loop mechanism in the weights, requiring its usage at inference time. The R^3-CNN architecture is able to surpass the recently proposed HTC model, while reducing the number of parameters significantly. Experiments on COCO minival 2017 dataset show performance boost independently from the utilized baseline model. The code is available online at https://github.com/IMPLabUniPr/mmdetection/tree/r3_cnn.

PDF Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Object Detection COCO minival R3-CNN (ResNet-50-FPN, DCN) box AP 44.8 # 112
AP50 64.3 # 40
AP75 48.9 # 33
APS 26.6 # 36
APM 48.3 # 25
APL 59.6 # 31
Instance Segmentation COCO minival R3-CNN (ResNet-50-FPN, GC-Net) mask AP 40.2 # 71
AP50 61.1 # 12
AP75 43.5 # 11
APM 42.8 # 6
APS 22.6 # 5
Object Detection COCO minival R3-CNN (ResNet-50-FPN, GC-Net) box AP 44.3 # 121
AP50 64.1 # 43
AP75 48.4 # 37
APS 27 # 30
APM 47.1 # 36
APL 58.9 # 36
Instance Segmentation COCO minival R3-CNN (ResNet-50-FPN, DCN) mask AP 40.4 # 69
AP50 61.3 # 11
AP75 44 # 10
APL 56.1 # 5
APM 43.6 # 5
APS 22.3 # 6
Instance Segmentation COCO minival R3-CNN (ResNet-50-FPN, GRoIE) mask AP 39.1 # 74
AP50 58.8 # 15
AP75 42.3 # 13
APL 54.3 # 7
APM 42.1 # 7
APS 20.7 # 7
Object Detection COCO minival R3-CNN (ResNet-50-FPN, GRoIE) AP50 61.2 # 68
AP75 45.6 # 58
APS 24.4 # 54
Instance Segmentation COCO minival R3-CNN (ResNet-50-FPN) mask AP 38.2 # 77
AP50 58 # 16
AP75 41.4 # 14
APL 52.8 # 10
APM 41 # 9
APS 20.4 # 8
Object Detection COCO minival R3-CNN (ResNet-50-FPN) box AP 42 # 144
AP50 61 # 71
AP75 46.3 # 52
APS 24.5 # 53
APM 45.2 # 51
APL 55.7 # 57
Instance Segmentation coco minval R3-CNN (ResNet-50-FPN, GC-Net) APL 56 # 1

Methods