--- kind: concept slug: coordinate-classification title: Coordinate classification summary: >- Predicting each keypoint as one distribution over discrete bins per image axis, instead of a 2-D heat map. The paper's variant (SimCC-style head with one gated attention unit) runs fast on CPU, GPU and mobile at 256x192 input. Main trade-off: a discrete grid and a soft-label loss to tune, against heat-map upsampling cost. One card so far; author-reported. tags: [pose-estimation, keypoints, real-time, simcc] papers: ["2303.07399"] related: [unsupervised-keypoint-detection, self-attention, efficient-vision-backbone] status: draft updated_at: 2026-10-24 --- ## Overview Keypoint methods usually output a heat map per joint and take its peak. A coordinate-classification method instead predicts two 1-D distributions per joint, one over the horizontal bins and one over the vertical bins. The joint position is the argmax of each distribution [arxiv:2303.07399 §3.1] [§3.4]. Training uses a Gaussian soft label over the bins around the true coordinate, and a KL divergence against the predicted distribution. The loss applies a temperature to both the outputs and the labels, and sets the label width per axis from the number of bins [arxiv:2303.07399 §3.4] [eq. 4] [eq. 5] [eq. 6]. Since each keypoint yields two vectors rather than a map, the output stage is small, which is the basis of the paper's speed claim [arxiv:2303.07399 §3.1] [§4.3]. ## Variants Only one card lists this slug, so the variants below are design choices the card's paper evaluates, not a survey of the field. The paper's development chain gives the COCO val AP of each step [arxiv:2303.07399 §3.3] [§3.4]. ### Full head: 7x7 convolution, GAU, temperature (RTMPose) Each keypoint gets one map from a 7x7 convolution. The map is flattened, projected to 256 dimensions and passed through one gated attention unit (GAU). Two linear layers give the x and y logits. The paper uses a temperature tau of 0.1 in the loss and a separate sigma for each axis [arxiv:2303.07399 §3.3] [§3.4]. Reported: 73.3 AP on COCO val after the 7x7 convolution, from a 72.8 step before it [arxiv:2303.07399 §3.4] [tab. 2]. Card: [RTMPose: Real-Time Multi-Person Pose Estimation based on MMPose](/papers/2303.07399). ### Head without the GAU block The paper's chain has a step before the GAU, at feature dimension 256, that reaches 71.4 AP on COCO val. Adding the GAU gives 71.9 in the same chain [arxiv:2303.07399 §3.3]. The gain is small and comes from a cumulative sequence of changes, not from an isolated ablation. Card: [RTMPose](/papers/2303.07399). ## When to use which The paper's speed tables are batch-1 timings on three setups: a Snapdragon 865 with ncnn, an Intel i7-11700 with ONNX Runtime on one thread, and a GTX 1660 Ti with TensorRT FP16 [arxiv:2303.07399 §4.3]. The timings are for the pose model alone unless marked as a pipeline. Numbers are author-reported. | Situation | Use | Why | |---|---|---| | Tight CPU or mobile budget, lower accuracy is acceptable | RTMPose-t or RTMPose-s | 9.02 ms and 13.89 ms on the Snapdragon 865 in FP16 [arxiv:2303.07399 tab. 10]. 65.8 and 69.6 COCO AP with Faster R-CNN and no extra data [tab. 4] | | Accuracy with a GPU or desktop CPU | RTMPose-m | 2.29 ms on the GTX 1660 Ti (TensorRT FP16) and 11.06 ms on the CPU, pose only [tab. 10] [tab. 11]. 73.6 AP with no extra data, 75.7 with AIC data and an RTMDet-m detector [tab. 4] | | Largest model in the paper's COCO table, latency is less of a concern | RTMPose-l | 4.16 GFLOPs, 18.85 ms on the CPU [tab. 10]. 76.6 AP with AIC data and an RTMDet-m detector [tab. 4] | | Compared with a heat-map model on the same hardware | RTMPose-m against HRNet-w32+UDP | 11.06 ms against 37.73 ms on the CPU, and 2.29 ms against 5.13 ms on the GPU [tab. 10]. The paper's AP comparison is in its own tables, not repeated here | | Whole-body keypoints, more compute is acceptable | RTMPose-l at 384x288 | 64.8 whole-body AP at 10.1 GFLOPs, against 61.1 at 4.5 GFLOPs for 256x192, with flip test [tab. 6] | | New domain with little labelled data | Start from AIC and COCO pre-training | RTMPose-m on AP-10K rises from 68.4 with ImageNet pre-training to 72.2 [tab. 8]. On CrowdPose and MPII the gains are 3.7 and 1.8 points [tab. 9] | Detector cost is not in the pose-model numbers. With RTMDet-nano at 320x320 as the detector, the RTMPose-m pipeline takes 26.61 ms on the CPU and 4.31 ms on the GPU in the paper's setup [arxiv:2303.07399 tab. 12]. Budget the detector separately. ## Implementation sketch The common core is the same for all variants: one distribution per axis, a soft target around the true coordinate, a KL loss, and an argmax at test time. The head shape and the bin count depend on the input size, which is 256x192 in the paper's configs [arxiv:2303.07399 §3.1]. ```pseudocode # Target for one visible keypoint, per axis [arxiv:2303.07399 §3.4] [eq. 4] # Input: keypoint (x, y) in crop pixels; sigma per axis [eq. 6] # Output: soft targets over the bins of each axis def soft_target(x, y, sigma_x, sigma_y, bins_x, bins_y, ratio): # Map pixels to bins; the official code uses 2 bins per pixel, # which the paper does not state mu_x, mu_y = x * ratio, y * ratio # Gaussian centred on the true bin, unnormalised [§3.4] [eq. 4] t_x = exp(-(bins_x - mu_x) ** 2 / (2 * sigma_x ** 2)) t_y = exp(-(bins_y - mu_y) ** 2 / (2 * sigma_y ** 2)) return t_x, t_y ``` ```pseudocode # Loss on temperature-softened distributions [arxiv:2303.07399 §3.4] [eq. 5] # Input: logits and soft targets per axis; tau = 0.1 # Output: scalar loss def coord_loss(logits, targets, tau=0.1): # Softmax over bins, applied to both model output and target [eq. 5] log_p = log_softmax(logits / tau) q = softmax(targets / tau) # KL divergence over the bins; the official code averages instead return sum(q * (log(q) - log_p)) ``` ```pseudocode # Decoding [arxiv:2303.07399 §3.1] [§3.4] # Input: logits_x (K, bins_x), logits_y (K, bins_y) # Output: keypoint (x, y) in crop pixels def decode(logits_x, logits_y, ratio): # Argmax per axis, then undo the bin-to-pixel ratio x = argmax(logits_x) / ratio y = argmax(logits_y) / ratio return x, y ``` ## Contradictions No cross-paper contradictions: only one card lists this slug. Within the card's paper, the abstract and the tables differ. The abstract gives 75.8 AP for RTMPose-m with flip test, and no table row has that value. The best table row for RTMPose-m is 75.7 [arxiv:2303.07399 §1] [tab. 4]. The abstract gives 72.2 AP for RTMPose-s. Table 10 gives 72.8 AP(GT) for the same model, and the paper does not define the GT column [arxiv:2303.07399 §1] [tab. 10]. ## Papers - [RTMPose: Real-Time Multi-Person Pose Estimation based on MMPose](/papers/2303.07399): a top-down pose estimator that predicts each keypoint as two 1-D SimCC distributions, with a CSPNeXt backbone and one GAU head.