BlurBall: Joint Ball & Blur Detection

• BlurBall is defined as a detection framework that jointly estimates ball position and blur attributes from multi-frame inputs, treating blur as a dynamic signal.
• It leverages an HRNet-based WASB model with Squeeze-and-Excitation channel attention and blur-aware heatmaps to boost detection accuracy and trajectory prediction.
• Empirical evaluations report significant F1 improvements and reduced blur estimation errors, demonstrating its effectiveness for fast-moving objects in sports analytics.

Searching arXiv for the cited BlurBall-related papers and closely related work. BlurBall is a table tennis ball detection framework that jointly estimates ball position and motion blur attributes from multi-frame inputs. It was introduced as a detector built on an HRNet-based baseline (WASB), with blur-aware heatmaps and lightweight channel attention through Squeeze-and-Excitation, and it is explicitly motivated by the observation that fast balls in racket sports often appear as streaks rather than distinct points (Gossard et al., 22 Sep 2025). In the supplied literature, the term also has adjacent uses: as a practical system description for intra-frame tracking by deblatting, based on "Intra-frame Object Tracking by Deblatting" (Kotera et al., 2019), and, in a different fluid-mechanical context, as a label for the stable suspension of a light ball in a slightly inclined liquid jet described in "Fluid juggling" (Soto et al., 2013). The dominant technical meaning, however, is the 2025 vision model for joint ball and blur estimation.

1. Definition, scope, and naming

BlurBall is defined as a detector that jointly estimates ball position and motion blur attributes from multi-frame inputs. Its central design choice is to place the annotation at the center of the blur streak rather than at the leading edge, and to supervise blur attributes explicitly through blur-aware heatmaps (Gossard et al., 22 Sep 2025). The model is described as suitable for table tennis ball tracking, with the broader claim that leveraging blur improves detection accuracy and enables more reliable trajectory prediction for real-time sports analytics.

The rationale for center-of-blur labeling is stated in geometric and learning-theoretic terms. Leading-edge labels break symmetry and create ambiguity about which end is “front,” whereas center-of-blur labels yield symmetric targets around the true location and therefore better align with Gaussian or soft heatmap formulations. The supplied description further states that blur carries velocity cues in both direction and magnitude, so a labeling convention that preserves the full streak is not merely an annotation preference but part of the measurement model itself (Gossard et al., 22 Sep 2025).

A recurrent misconception is to treat motion blur only as a nuisance variable to be suppressed by deblurring or by shorter exposure. The BlurBall formulation instead treats blur as signal: the spatial support of the heatmap spans the blur region, and blur orientation and length are decoded from the predicted blob. This suggests a shift from point detection under blur to structured estimation from blur.

2. Labeling convention and dataset

The annotation schema places the ball center at the midpoint of the blur streak, denoted $\mathbf{p}_b = (x_b, y_b)$, and represents blur by a straight line segment with angle $\theta$ and half-length $l$ (Gossard et al., 22 Sep 2025). The extremities are

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$

and

$$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$

The dataset described for BlurBall contains 64,119 frames total from 26 different recordings with varied POV, table colors, lighting, and static cameras. The train split comprises 22 games, 363 clips, and 51,423 frames with blur ratio $\approx 0.58$, while the test split comprises 4 games, 80 clips, and 12,696 frames with blur ratio $\approx 0.77$ (Gossard et al., 22 Sep 2025). Blur prevalence is reported as approximately $62\%$ of frames, and the blur half-length distribution is described as most common around approximately $5$ px, with a maximum $l = 73$ px.

The CSV format is given as columns Frame, Visibility, X, Y, [](https://www.emergentmind.com/topics/free-lm-framework)θ, l, with angles in degrees, $\theta$0 in pixels, and $\theta$1 in half-length pixels. Camera calibration is provided per match through focal length $\theta$2 and camera extrinsics $\theta$3, while exposure times are not reported and are instead estimated in downstream trajectory fitting (Gossard et al., 22 Sep 2025).

The supplied labeling guidelines impose a straight-line assumption for blur during exposure. Rare frames at bounce may violate this through non-convex blobs, and angle estimation is described as unreliable for very short blur, specifically for cases such as $\theta$4 px. For that reason, angle MAE is computed only for $\theta$5 px. Traditional leading-edge labels are also provided for comparison experiments, allowing direct measurement of the effect of annotation convention.

3. Motion blur model and blur-aware estimation

The BlurBall formulation uses a time-integration model of motion blur under constant image-plane velocity:

$\theta$6

Under uniform exposure over time $\theta$7 and constant image-plane velocity $\theta$8, this is equivalent to convolution with a line PSF aligned to motion:

$\theta$9

The relation between blur length and velocity is given as

$l$0

with per-axis components

$l$1

These equations formalize why blur can be used as a velocity cue rather than treated purely as degradation (Gossard et al., 22 Sep 2025).

In the model parameterization, each frame contributes $l$2, $l$3, full blur length $l$4, and a confidence value. Width $l$5 and opacity or intensity $l$6 are explicitly stated as not annotated in the dataset and not estimated explicitly by the model. Angle periodicity is handled through PCA decoding of the principal axis, so angle comparisons should use the shortest angular difference modulo $l$7 because the principal axis has no inherent direction.

The blur-aware heatmap departs from standard center-point supervision. In the supplied description, standard WASB ground-truth heatmaps label a disk around the center point, whereas BlurBall redefines the target to cover the entire blur segment thickened by radius $l$8. The real-valued HLSM target is defined over pixels within distance $l$9 of the segment by

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$0

with $$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$1 chosen so the non-zero minimum equals $$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$2 (Gossard et al., 22 Sep 2025). This gives the detector a supervision target whose support matches the observed streak.

4. Architecture, decoding, and training

BlurBall uses an HRNet-based WASB baseline with multiple-frame input and MIMO-style inference, where frames are stacked as channels. The standard setting uses windows of 3 frames, and “steps=1” versus “steps=3” refers to the stride with which those windows are advanced (Gossard et al., 22 Sep 2025). Channel-wise attention is introduced through Squeeze-and-Excitation, defined by

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$3

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$4

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$5

The supplied account states that SE provided the most consistent improvements among ECA, CA, and SE.

Position and blur are jointly predicted from a single heatmap output. Position is computed as the weighted mean of heatmap values. Blur orientation is obtained by PCA on the thresholded blob, with principal axis $$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$6 and

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$7

Blur half-length is then decoded by projection of blob pixels onto $$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$8:

$$\mathbf{p}_1 = \mathbf{p}_b + (l \cos \theta, l \sin \theta)$$9

The default threshold for blob extraction is $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$0, while $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$1 is reported as tuned for 1-step inference to optimize F1 (Gossard et al., 22 Sep 2025).

Training uses quality focal loss on the real-valued blur-aware heatmap:

$$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$2

No explicit blur regression loss is applied; $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$3 and $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$4 are decoded from the heatmap rather than produced by separate heads. The reported training setup uses Adam, 30 epochs, batch size 8, input size $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$5, HLSM starting at epoch 20, distance threshold $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$6, and $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$7, with training from scratch on an NVIDIA GeForce RTX 2080 Ti (Gossard et al., 22 Sep 2025). Pretraining on badminton is reported not to have helped.

5. Evaluation, ablations, and trajectory prediction

The evaluation protocol includes detection metrics—F1, Accuracy, Average Precision, Recall, Precision, FPS, and parameter count—and blur estimation metrics—MAE for blur length $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$8 in pixels and angle $$\mathbf{p}_2 = \mathbf{p}_b - (l \cos \theta, l \sin \theta).$$9 in degrees (Gossard et al., 22 Sep 2025). Across many detectors, midpoint labeling improves detection relative to traditional front labeling. Representative F1 changes reported in the supplied description include DeepBall $\approx 0.58$0, DeepBall-Large $\approx 0.58$1, BallSeg $\approx 0.58$2, TrackNetV2 $\approx 0.58$3, Monotrack $\approx 0.58$4, WASB (steps=3) $\approx 0.58$5, and WASB (steps=1) $\approx 0.58$6 (Gossard et al., 22 Sep 2025).

Ablation results attribute much of the gain to blur-aware supervision. For the WASB baseline using blur estimation via background subtraction, the reported blur errors are $\approx 0.58$7 MAE $\approx 0.58$8 px (1-step) and $\approx 0.58$9 px (3-step), with $\approx 0.77$0 MAE $\approx 0.77$1 (1-step) and $\approx 0.77$2 (3-step). Adding blur-aware labels without attention reduces these to $\approx 0.77$3 MAE $\approx 0.77$4 px (1-step) and $\approx 0.77$5 px (3-step), and $\approx 0.77$6 MAE $\approx 0.77$7 (1-step) and $\approx 0.77$8 (3-step), while improving detection to F1 $\approx 0.77$9 and $62\%$0, with AP $62\%$1 and $62\%$2. The SE-attention variant designated BlurBall is reported as the best overall configuration, with F1 $62\%$3 (1-step) and $62\%$4 (3-step), Acc $62\%$5 and $62\%$6, AP $62\%$7 and $62\%$8, $62\%$9 MAE $5$0 px and $5$1 px, $5$2 MAE $5$3 and $5$4, parameter count approximately $5$5M, and FPS approximately $5$6 in the 3-step table column (Gossard et al., 22 Sep 2025). With threshold tuning to $5$7 for 1-step inference, F1 rises to $5$8 and Accuracy to $5$9, with a slight AP decrease and FPS approximately $l = 73$0.

The same blur estimates are then used for trajectory prediction. Using pixel scale $l = 73$1 and exposure $l = 73$2, the velocity magnitude is given by

$l = 73$3

The reported trajectory fitting uses quadratic polynomials $l = 73$4 and $l = 73$5 under a constant-forces assumption, with image-plane velocity constraints derived from $l = 73$6 and $l = 73$7. The objective functions are given in the supplied synthesis as

$l = 73$8

and

$l = 73$9

When fit from only 3 frames, the reported prediction error improves from MAE $\theta$00 px for a position-only fit to MAE $\theta$01 px for a position-plus-blur fit (Gossard et al., 22 Sep 2025). This supports the interpretation that blur is an informative dynamic measurement.

6. Relation to deblatting and other uses of the term

BlurBall is closely related in spirit to earlier work on intra-frame trajectory recovery from motion blur. "Intra-frame Object Tracking by Deblatting" formulates a forward model

$\theta$02

where $\theta$03 is a motion PSF supported on the trajectory manifold, $\theta$04 is the sharp appearance, $\theta$05 is the relaxed mask, and $\theta$06 is the background (Kotera et al., 2019). That framework estimates blur by solving intertwined inverse problems, blind deblurring and image matting, termed deblatting, and then fits a piecewise quadratic curve to the recovered PSF. Its dataset uses the Trajectory-IoU metric, and the reported averages are CSR-DCF with TIoU $\theta$07 and Recall $\theta$08, FMO with TIoU $\theta$09 and Recall $\theta$10, TbD with template and $\theta$11 at TIoU $\theta$12 and Recall $\theta$13, and TbD-Oracle at TIoU $\theta$14 (Kotera et al., 2019).

The conceptual affinity is direct. Both formulations treat the blur trace as evidence of within-frame motion rather than as a corruption to be removed. The deblatting framework is more explicitly inverse-problem-based, with ADMM updates over $\theta$15, $\theta$16, and $\theta$17, while BlurBall decodes blur attributes directly from a supervised heatmap. A plausible implication is that BlurBall occupies a lower-overhead point in the design space, whereas deblatting aims at a fuller physical reconstruction of the blur process.

The term also appears in a distinct fluid-dynamics usage in the supplied synthesis of "Fluid juggling," where “BlurBall” denotes the stable suspension of a light ball within a slightly inclined liquid jet (Soto et al., 2013). In that description, the ball is trapped because the impinging jet forms a thinner, faster annular film, with a Bernoulli-like pressure reduction and momentum-flux support against gravity. The key relations are presented as

$\theta$18

and

$\theta$19

This is not a computer-vision usage, but it is useful for disambiguation because the same label has been applied to both a blur-aware tracking concept and a hydrodynamic suspension phenomenon. The two topics are unrelated except for the naming coincidence.

7. Limitations, applicability, and future directions

The supplied description identifies BlurBall as especially suitable for other racket sports, such as badminton and tennis, and more generally for fast-moving objects with approximately linear blur during exposure (Gossard et al., 22 Sep 2025). Its main assumptions are linear motion during exposure, a single streak approximated by a straight segment, and conditions under which blur can be decoded reliably from the predicted heatmap. Static cameras are described as favorable, while dynamic cameras can reduce frame-differencing efficacy and affect blur decoding.

Failure modes listed in the supplied text include severe occlusions, complex backgrounds, white distractors such as hands, logos, or shoes, camera shake, and rolling-shutter artifacts that can bend streaks (Gossard et al., 22 Sep 2025). Absolute velocity estimation also depends on exposure metadata $\theta$20 and pixel scale $\theta$21; when those are absent, the model can still provide relative velocity cues from $\theta$22 and $\theta$23, but not fully calibrated speed. This suggests that the method is strongest when geometric calibration and approximate exposure conditions are available.

The reported future directions include end-to-end deblurring plus detection, physics-informed priors involving drag and Magnus effect, better temporal modeling through GRU or transformers, explicit blur heads with angular wrap-around loss, robust handling of rolling shutter, scene context for distractor suppression, and integration with 3D reconstruction and spin estimation pipelines (Gossard et al., 22 Sep 2025). Given the relation to deblatting, a plausible implication is that future systems may combine supervised blur-aware detection with inverse-problem structure, allowing both the efficiency of heatmap-based inference and the identifiability benefits of trajectory-constrained PSF estimation.