To boost or not to boost? On the limits of boosted trees for object detection

Abstract

Abstract-We aim to study the modeling limitations of the commonly employed boosted decision trees classifier. Inspired by the success of large, data-hungry visual recognition models (e.g. deep convolutional neural networks), this paper focuses on the relationship between modeling capacity of the weak learners, dataset size, and dataset properties. A set of novel experiments on the Caltech Pedestrian Detection benchmark results in the best known performance among non-CNN techniques while operating at fast run-time speed. Furthermore, the performance is on par with deep architectures (9.71% log-average miss rate), while using only HOG+LUV channels as features. The conclusions from this study are shown to generalize over different object detection domains as demonstrated on the FDDB face detection benchmark (93.37% accuracy). Despite the impressive performance, this study reveals the limited modeling capacity of the common boosted trees model, motivating a need for architectural changes in order to compete with multi-level and very deep architectures.

Occlusion-Aware R-CNN: Detecting Pedestrians in a Crowd

“…We directly fine-tune the detection models pre-trained on CityPersons [67] of the proposed OR-CNN method on the training set in Caltech-USA. Similar to [55], we evaluate the OR-CNN method on the Reasonable subset of the Caltech-USA dataset, and compare it to other state-of-the-art methods (e.g., [55,66,50,5,6,63,30,25,49,9,34,15]) in Figure 4. Notably, the Reasonable subset (occlusion < 35%) only includes the pedestrians with at least 50 pixels tall, which is widely used to evaluate the pedestrian detectors.…”

SFA: Small Faces Attention Face Detector

“…Our method is trained on the training set of the WIDER FACE dataset and evaluate on its validation and testing set against the recently published state-of-theart face detection methods including Zhu et al [27], S 3 FD [21], SSH [20], HR [26], MSCNN [33], CMS-RCNN [24], Multitask Cascade CNN [18], LDCF+ [34] and Multiscale Cascade CNN [30]. The precision-recall curves and mAP values on WIDER FACE validation and testing sets are presented in Fig.…”

“…In these datasets, we resize the shortest side of the input images to 400 pixels while keeping the larger side less than 800 pixels, leading to an inference speed of more than 20 FPS. And we directly use our final SFA detector model in Experiment XIII and compare SFA against the recently published state-of-the-art methods including FD-CNN [35], ICC-CNN [36], RSA [37], S 3 FD [21], FaceBoxes [22], HR [26], HR-ER [26], DeepIR [38], LDCF+ [34], UnitBox [39], Conv3D [40], Faster RCNN [41] and MTCNN [18] on FDDB dataset. For a more fair comparison, the predicted bounding boxes are converted to bounding ellipses.…”

Face-MagNet: Magnifying Feature Maps to Detect Small Faces

“…We use the same network that is trained on WIDER training set without any further fine-tuning. There are many [7], Faster-RCNN [10,21], Multiscale Cascade CNN [25], Faceness-WIDER [22], HyperFace [20], HeadHunter [15], LDCF [18], Conv3D [12]. Some qualitative detection results from all of the datasets are presented in Figure 10.…”