When a person occupies only 32×32 or 64×48 pixels, the parts of the body vanish at different rates. The torso still covers many pixels. A wrist or ankle may cover less than one cell of the network's coarse feature maps. Current low-resolution pose estimators handle this by preserving more pixels during downsampling or by refining coordinates below pixel precision. Both still predict every joint independently, so an ankle with no visual evidence is guessed with no reference to the knee above it. This proposal tests a specific remedy. The network's decoder produces feature maps at several resolutions, from coarse to fine. We assign each depth of the body's skeleton tree to one of these resolutions: torso at the coarsest map, then shoulders and hips, then elbows and knees, and finally wrists and ankles at the finest map. Each child joint is predicted as a short offset from its already-predicted parent. Every stage receives its own loss on absolute joint positions, a practice called deep supervision.
The plain-language claim is this. If each part of the body is learned at the resolution where it is still visible, and each finer part is anchored to a coarser part, then the small extremities that low resolution erases can be recovered from where their parents are. This works better than recovering them from local pixels that no longer exist, and better than cleaning up independently predicted coordinates afterwards. The test is narrow and checkable. Gains should concentrate on distal joints, grow as input resolution shrinks, and disappear or reverse when the order of the tree is scrambled across scales.
Core Claim
Three bodies of work each hold one piece of this. None holds the combination. Low-resolution pose methods show the failure clearly. Information-lossless downsampling (MGIL), which moves pixels into channels instead of discarding them, plus dilated context, reaches 37.5 AP on COCO val2017 at 32×32 with HRNet-W32 (Focus on Low-Resolution Information). Confidence-aware coupling of heatmaps and offsets (CAL) reaches 72.3% AP at 128×96 (Low-resolution Human Pose Estimation). The MGIL authors argue that skeletal context becomes the main cue when local texture is gone. Yet both methods treat joints as independent output channels and encode the skeleton only implicitly, through receptive fields. These papers motivate the proposal and supply the strongest baselines.
Kinematic regression shows that anchoring children to parents reduces variance at the extremities. Sun et al. regress the palm first and fingers relative to it (Cascaded hand pose regression). Deep Kinematic Pose Regression chains rotations through a forward-kinematics layer (Deep Kinematic Pose Regression). That paper also supplies a design rule we reuse: put the loss on composed Cartesian positions, not on relative parameters, because small proximal errors are amplified at the extremities. HiPART orders generation from center to periphery (HiPART), but it lifts 2D coordinates from an external detector and never reads pixels.
Deep supervision shows that intermediate losses help pose refinement. The Stacked Hourglass applies the same full-skeleton heatmap loss after every module (Stacked Hourglass Networks). DISCO supervises side branches with an ordered chain of semantic concepts (Deep Supervision with Intermediate Concepts).
What is new here is the alignment of two hierarchies inside one image-to-pose network. One is the hierarchy of feature resolution in the decoder. The other is the hierarchy of the skeleton tree. Children read image features at their parent's predicted location and are supervised in absolute coordinates at every stage. No work in the file supervises intermediate stages along the branches of a kinematic tree. No work tests such a structure under resolution degradation.
Research Program
The model keeps a standard backbone, HRNet-W32, so that comparisons with MGIL are direct. It adds a decoder with three or four output scales. At the coarsest scale, a head predicts the root (pelvis/torso center, or the midpoint of hips and shoulders for COCO's 17-joint layout) using a heatmap plus offset.
At each finer scale $s$, every joint $j$ at tree depth $s$ is predicted relative to its parent $p(j)$. The head bilinearly samples features $F_s$ at the parent's current estimate $\hat x_{p(j)}$. It then regresses a 2D bone vector $b_j$, whose length is normalized by the predicted torso size. The joint position is
$$\hat x_j = \hat x_{p(j)} + b_j\big(F_s(\hat x_{p(j)})\big) + r_j,$$
where $r_j$ is a small residual. It comes from a local heatmap at scale $s$ and lets visible evidence override the prior. Without this residual, errors would compound down the chain. The loss at every stage is placed on the absolute positions $\hat x_j$ of all joints resolved so far, weighted by COCO's per-joint OKS falloff. Following Deep Kinematic Pose Regression, gradients therefore flow from wrist error back into the elbow and shoulder estimates. To limit the over-constraint risk discussed below, the intermediate losses are annealed. Their weight starts equal to the final loss and decays to a small value over training.
The analysis has two parts. First, we measure each joint's effective pixel support: its OKS sigma times the person's scale, divided by the feature stride. We plot per-joint error against support for every method. The hypothesis predicts that the advantage appears exactly where support falls below about one feature cell. Second, we measure structural plausibility directly, as the variance across the test set of limb-length ratios and the rate of left/right limb swaps.
Baselines
The naive reference is HRNet-W32 trained and tested at the degraded resolution with a standard heatmap head and with SimCC. It shows how much any method helps. On COCO at 32×32 its level is implied by MGIL's reported 7.7-point gain over SimCC.
The strongest existing method at 32×32 is MGIL with HRNet-W32, reported at 37.5 AP on COCO val2017 (Focus on Low-Resolution Information). At 128×96 the strongest is CAL, reported at 72.3% AP (Low-resolution Human Pose Estimation). At 64×48, CAL reports only a lead of more than 10 AP over its nearest competitor, with no absolute number given in the file. On MPII at 32×32, MGIL reports a gain of 6.388 [email protected], but no absolute value is found on Lacuna. We will retrain both MGIL and CAL in our codebase, with the same data, augmentation, epochs, and hyperparameter search budget as our model, and report our reproduced numbers alongside the published ones.
Because our head is orthogonal to MGIL's downsampling module, the decisive comparison is MGIL with its standard head against MGIL with the kinematic head. Beating MGIL means exceeding its reproduced AP by more than seed variance across three seeds. The gain must come mostly from wrists, ankles, elbows, and knees, at matched compute (FLOPs within about 10%).
A Stacked-Hourglass-style control applies identical full-skeleton supervision at every decoder scale (Stacked Hourglass Networks). It separates "any deep supervision helps" from "tree-ordered supervision helps." The key ablation is a scrambled-hierarchy model with the same heads and losses but the depth-to-scale assignment reversed: extremities at the coarse scale, torso at the fine scale. A second ablation removes parent-conditioned feature sampling but keeps the tree-ordered losses.
Why not just…?
Why not just run a learned pose prior on the detector's output? One could clean up MGIL's coordinates with a skeleton denoiser in the spirit of HiPART's center-to-periphery generation (HiPART). This is the most serious threat. A coordinate-only refiner can fix implausible limbs but cannot look again at the image. Our claim is that sampling features at the parent location recovers faint distal evidence that independent heatmaps dilute in background clutter. We will train a transformer refiner, matched in parameters to our head, on MGIL outputs. If it matches our distal-joint AP, the in-network conditioning is not needed and the contribution reduces to a cheap post-processing result. We would report that honestly.
Why not just add a bone-length or limb-consistency loss to MGIL or CAL? This adds structure at almost no cost. However, it constrains only the output and does not change which features the extremity heads read. We include it as a baseline. If it closes most of the gap, the scale-assignment idea is unnecessary.
Why not just super-resolve the input first? CAL's authors report that processing low-resolution input directly beats upscaling with super-resolution, which adds artificial high-frequency detail (Low-resolution Human Pose Estimation). We therefore include super-resolution plus HRNet only as a sanity reference.
Why not worry that intermediate supervision makes the model brittle? Recurrent CNNs forced to imitate prescribed maze-search traces collapsed under topology shift (When Intermediate Supervision Doesn't Help). This is a real constraint on the design. Our targets are physical joint positions, not procedural traces, and the residual term plus loss annealing leaves the network free to depart from the chain. We still test brittleness directly. We train at 64×48 and evaluate at 32×32 and 128×96. We also evaluate on MPII's occlusion-flagged and extreme-torso-tilt subsets, which are poses where a rigid tree ordering could mislead.
Proposed Experiments
The main experiments use COCO keypoints (Microsoft COCO) and MPII (2D Human Pose Estimation benchmark; 2D Human Pose Estimation: New Benchmark and State of the Art Analysis), with person crops downsampled to 128×96, 64×48, and 32×32 following the protocols of CAL and MGIL. We report COCO AP broken down by joint group and MPII [email protected] per joint. A secondary experiment downsamples H3WB RGB crops (H3WB) to probe hands and feet, where pixel support is smallest. The only RGB-to-3D reference there is the 132.6 mm MPJPE of a two-stage approach at full resolution, and no low-resolution H3WB result is found on Lacuna.
Three outcomes would decide the question. First, the kinematic head added to MGIL should beat MGIL alone at 32×32, with at least half the gain on distal joints. Second, the gain should shrink toward zero at full resolution. Third, the scrambled-hierarchy ablation should lose most of the gain. If tree-ordered and scrambled supervision perform the same, alignment of resolution with the skeleton is not the mechanism. If the coordinate-only refiner matches us, the image-conditioned design is not justified. Either result would be reported as the finding.
References
- Focus on Low-Resolution Information: Multi-Granular Information-Lossless Model for Low-Resolution Human Pose Estimation
- Low-resolution human pose estimation
- Cascaded hand pose regression
- Deep Kinematic Pose Regression
- HiPART: Hierarchical Pose AutoRegressive Transformer for Occluded 3D Human Pose Estimation
- Stacked Hourglass Networks for Human Pose Estimation
- Deep Supervision with Intermediate Concepts
- When Intermediate Supervision Doesn’t Help: Evidence from Recurrent CNNs
- Microsoft COCO: Common Objects in Context
- 2D Human Pose Estimation: New Benchmark and State of the Art Analysis
- H3WB: Human3.6M 3D WholeBody Dataset and Benchmark