Look Closer to Segment Better:
Boundary Patch Refinement for Instance Segmentation

Given an instance mask produced by an instance segmentation model, we first need to determine which part of the mask should be refined. Based on the findings of previous works [9, 50] and our verification experiments in Table 1, we propose an effective sliding-window style algorithm to extract a series of patches along the predicted instance boundaries. Specifically, we densely assign a group of squared bounding boxes where the central areas of the box should cover the boundary pixels, shown in Figure 2(b). The obtained boxes still contain large overlaps and redundancies, thus we further apply a Non-Maximum Suppression (NMS) algorithm to filter out a subset of patches (Figure 2c). Empirically, with the larger overlaps, the segmentation performance can be boosted, while simultaneously suffering from the larger computational cost. We can adjust the NMS threshold to control the amount of overlap to achieve a better speed/accuracy trade-off. In addition to image patches, we also extract the corresponding binary mask patches from the given instance mask. The concatenated image and mask patches (Figures 2d and 2e) are resized and fed into the following boundary patch refinement network.

Mask Patch. The benefit of the binary mask patch is that it accelerates training convergence and provides location guidance for the instance to be segmented. As discussed in the previous works on semantic segmentation [41, 49], context information plays a vital role for pixel-wise classification. Therefore, the cropped image patches are hard to be classified independently due to the limited context information. With the help of location and semantic information provided by the mask patches, the refinement network can eliminate the need for learning instance-level semantics from scratch. Instead, the refinement network only needs to learn how to locate the hard pixels around the decision boundary and push them to the correct side. We believe this goal can be achieved by exploring low-level image properties (\egcolour consistency and contrast) provided in the local and high-resolution image patches. More importantly, the adjacent instances are likely to share an identical boundary patch, while the learning goals are totally different and ambiguous. Together with different mask patches for each instance, these issues can be avoided. As compared in Table 2, the model has trouble to converge without using the mask patches, examples of which are shown in Figure 3.

Boundary Patch Refinement Network. The role of this refinement network is to perform binary segmentation for each extracted boundary patch individually. Any semantic segmentation model can be employed for this task by simply modifying the input channels to 4 (3 for the RGB image patch and 1 for the binary mask patch) and output classes to 2. For the sake of convenience, we adopt the state-of-the-art HRNetV2 [41] as the refinement network in our implementation, which can maintain high-resolution representation throughout the whole network. By increasing the input size appropriately, the boundary patches can be processed with much higher resolution than in previous methods.

Reassembling. The refined boundary patches are reassembled into a compact instance-level mask by replacing their previous predictions. Predictions are unchanged for those pixels without refinement. For the overlapping areas of adjacent patches, the results are aggregated by simply averaging the output logits and applying a threshold of 0.5 to distinguish the foreground and background.

The refinement network is trained based on the boundary patches extracted from training images and tested on validation or testing images. We do not directly train or fine-tune the instance segmentation models. During training, we only extract boundary patches from instances whose predicted masks have an Intersection over Union (IoU) overlap larger than 0.5 with the ground-truth masks, while all predicted instances are retained during inference. The model outputs are supervised with the corresponding ground-truth mask patches using the pixel-wise binary cross-entropy loss. We simply fix the NMS eliminating threshold to 0.25 during training, while adopting different thresholds during inference based on the speed requirements. See Appendix A for more implementation details.

Datasets. We mainly report the results on Cityscapes [10], a real-world dataset with high-quality instance segmentation annotations. We only used the fine data, containing $2,975/500/1,525$ images for train/val/test, which are collected from 27 cities, with a high resolution of 1024$\times$2048. Eight instance categories are involved, including bicycle, bus, person, train, truck, motorcycle, car, and rider.

Metrics. The COCO-style [25] mask AP (averaged over 10 IoU thresholds ranging from 0.5 to 0.95 in the step of 0.05), AP₅₀ (AP at an IoU of 0.5) and AP_S/AP_M/AP_L (for small/medium/large instances) were reported in most of our experiments. The official Cityscapes-style AP [10] was only used to report the final results for a fair comparison, which is slightly higher than the COCO-style AP. Similar to [23, 37, 50], we also used a boundary F-score to evaluate the quality of the predicted boundaries. A mask was considered correct if the boundary is within a certain distance threshold from the ground-truth. We used a threshold of 1px and only compute for true positives that are determined on the same 10 IoU thresholds ranging from 0.5 to 0.95. The boundary F-score was computed in an instance-wise manner and then averaged over them, termed AF.

We investigated the effectiveness of the proposed framework through extensive ablation experiments on the configurable design choices. We started the refinement with the results of Mask R-CNN ResNet-FPN-50 baseline trained on Cityscapes fine data (with COCO pre-training). We adopted the lightweight HRNetV2-W18-Small as the refinement network in the default setting, with input size equal to 128$\times$128. The boundary patches were extracted with patch size equal to 64$\times$64 without padding, and the inference NMS threshold was set to 0.25 by default.

Effects of Mask Patch. To validate the effect of mask patch for boundary refinement, we made a comparison by eliminating the mask patches while keeping other settings unchanged. As indicated in Table 2, the model trained with image patches solely yielded a terrible result, even worse than the segmentation results before refinement. However, together with mask patches, we achieved a significant improvement (+$3.4\%$ in AP, +$11.9\%$ in AF) by refining the Mask R-CNN segmentation results. We also show some patch-wise examples in Figure 3. For a simple case with one dominant instance in the image patch (first row), both of the models (w/ or w/o mask patch) produced reasonable results. However, as for cases with multiple instances crowded in the image patch, the model without mask patch (last column) failed to distinguish which object should be segmented, leading to coarse (4th row) or completely wrong (2nd and 3rd rows) predictions. In contrast, with the help of mask patches, we produced high-quality predictions with accurate and distinct boundaries (3rd column).

Patch Size. We increased the boundary patch size by cropping with a larger box and/or with padding. Note that the padded areas were only used to enrich the context and not used for reassembling. As the patch size gets larger, the model becomes less focused but can access more context information. In Table 3, we compared various choices and found that the 64$\times$64 patch without padding works better. We used this setting in all experiments.

Different Patch Extraction Schemes. The most important contribution of this work is the idea of looking closer at instance boundaries to achieve better segmentation results. There are multiple choices about how to extract the boundary patches for refinement. We compared three extraction schemes and listed the results in Table 4. The most straightforward scheme is dividing the input image into a group of patches according to the pre-defined grids, and then picking only the patches that covering the predicted boundaries. We varied the patch size and found the results were consistently worse than our proposed “dense sampling + NMS filtering” scheme. One of the most important reasons is the imbalanced foreground/background ratio. We observed that some extracted patches are almost entirely filled with either foreground or background pixels. These patches are hard to refine due to the lack of context, thus leading to sub-optimal results. In contrast, by restricting the center of patches to cover the boundary pixels, the imbalance problem can be alleviated. Another scheme, similar to some previous works [23, 28], is cropping the whole instance based on the detected bounding box and further re-segmenting the instance patch. As shown, even though the input size was increased to 512$\times$512, the results are still sub-optimal, which demonstrated the effectiveness of our local boundary patches. See Appendix B for detailed descriptions.

Input Size of the Refinement Network. The extracted patches are resized into a larger scale before refinement. Table 5 shows the impact of input size. We also report the approximate inference speed of the refinement network, with a fixed batch size of 135 (on average 135 patches per image). As the input size increases, the AP and AF scores increase accordingly, and slightly drop after 256. The boundaries can be processed with the higher resolution with the larger input size, thus more details can be retained.

Alternatives of refinement network. We compared different backbones for our refinement network in Table 6. As shown, a stronger backbone usually lead to higher performance, but at the expense of lower speed. Since the model essentially performs binary segmentation for patches, it can further benefit from the advances in semantic segmentation, such as increasing the resolution of feature maps [6, 7, 41] and more effective backbones [45, 51].

NMS Eliminating Threshold. We studied the impact of different NMS eliminating thresholds during inference, shown in Table 7. As the threshold gets larger, the number of boundary patches increases rapidly. The overlap of adjacent patches provides a chance to correct unreliable predictions of the inferior patches. As shown, the resulting boundary quality was consistently improved with a larger threshold, and reached saturation around 0.55. We believe a better speed/accuracy trade-off can be achieved by setting a proper threshold.

What the BPR model learned is a general ability to correct error pixels around instance boundaries. We can easily transfer this ability of boundary refinement to refine the results of any instance segmentation model. Specifically, once we get a model trained on the boundary patches extracted from the train-set predictions of Mask R-CNN on Cityscapes, we can make inference to refine any predictions (on Cityscapes train/val/test sets) produced by any models (not only Mask R-CNN), without the need for training from scratch. After training, the BPR model becomes model-agnostic, similar to SegFix [50]. We validated the transferability by applying the model trained on Mask R-CNN results to refine the predictions of PointRend [19] and SegFix [50]. Note that these two methods are also designed to improve boundary quality in segmentation. As shown in Table 9, the transferred model still improved the results of PointRend and SegFix by a large margin, suggesting that our method is compatible with them.

Comparison with State-of-the-art Methods. We integrated the optimal design choices and hyperparameters found in above ablation experiments into a stronger BPR model. Specifically, we adopted the HRNetV2-W48 as our refinement network, with 256$\times$256 input patches resized from 64$\times$64, and a NMS threshold of 0.55 during inference. We evaluated the framework on Cityscapes val and test sets and compared the performance against some state-of-the-art methods in Table 8. (1) Compared with the Mask R-CNN baseline, we achieved a significant improvement (+$\mathbf{4.3\%}$ AP in both val and test). We outperformed SegFix [50] by a large margin, which is also a boundary refinement module applied to the same baseline with ours. Furthermore, by applying our BPR model to the results already refined by SegFix, we can still improved a lot (slightly lower than applying BPR only). (2) We transferred the above BPR model to refine the results of the stronger PolyTransform [23] baseline ($1^{st}$ place at CVPR 2020). Our “PolyTransform + BPR” consistently improved $2.3\%$ AP on Cityscapes test set and also outperformed “PolyTransform + SegFix” ($2^{nd}$ place at ECCV 2020) by a large margin (+$1.2\%$). By applying BPR to “PolyTransform + SegFix”, we established a new state-of-the-art on Cityscapes test with AP of 42.7%, reaching $\mathbf{1^{st}}$ place on the Cityscapes leaderboard by the CVPR 2021 submission deadline.

Qualitative Results. We show some qualitative results on Cityscapes val in Figure 4. Compared with the coarse predictions of Mask R-CNN, our BPR framework generated substantially better instance segmentation results with precise and distinct boundaries. It largely alleviated the over-smoothing issues [19] in previous methods caused by the low resolution feature maps. More results are included in Appendix E. In addition, we also provided a detailed limitation analysis in Appendix F.

Speed. Only the speed of refinement network was considered in Table 5 and 6, excluding the patch extraction and reassembling time. As a whole pipeline, it takes about 211ms to process a single Cityscapes image (1024$\times$2048) on a single RTX 2080Ti GPU under the default setting of ablation experiments, which is still much faster than PolyTransform [23]. The detailed speed calculation and more speed analysis are included in Appendix C.

Results on COCO Dataset. To demonstrate the generality of our framework, we also report the results on the more challenging COCO dataset [25], which contains 80 categories and more images (118k/5k for train/val). It is important to note that the coarse annotations in COCO may not fully reflect the improvements in mask quality [14]. Following PointRend [19], we further report the AP^⋆ measured using the higher quality LVIS [14] annotations. We randomly sampled about $8\%$ of instances for fast training. As shown in Table 10, we improved the powerful Mask R-CNN ResNeXt-FPN-101 baseline by $0.8\%$ AP and $1.7\%$ AP^⋆ on val2017. The coarse annotations on COCO train2017 may provide ambiguous optimization objectives, especially for our local boundary patches. It may mislead the learning of our BPR model, leading to suboptimal results. This issue was also observed in some contour-based instance segmentation methods [33, 44, 47]. We believe that training with more instances on higher quality annotations (\egLVIS) can further improve the results. More analysis on COCO dataset is included in Appendix D.