Systems that find when actions happen in long videos are often trained with only a label saying which actions appear somewhere in the video. This setting is called weakly supervised temporal action localization. Such systems fail in two opposite ways. They shrink onto the single most telling instant, the "smoking gun" frame, or they spill over into surrounding scenery that always accompanies the action, such as the pool around a dive. The standard score, mean average precision over overlap thresholds, mixes these two errors into one number, so a method that fixes one while worsening the other can look neutral. Nuclear physics faced the same measurement problem and solved it with the nuclear modification factor $R_{AA}$. This is the yield observed in a heavy-ion collision divided by the yield expected if the collision were a plain sum of independent proton–proton collisions. The ratio is reported in bins of collision "centrality," which measures how much nuclear matter the probe had to pass through. A value of $R=1$ means the medium was transparent. Values below or above 1 name the direction of the distortion.
We propose to borrow this measurement logic, not any physical mechanism. The plan has two parts. First, we build a coverage modification factor for action localizers: predicted action time divided by true action time, reported in bins of video "thickness" (length and background fraction), and split into an instance-count part and a per-instance span part. Second, we test a training-time correction. Current methods assume that the amount of foreground scales linearly with video length. We replace that assumption with a two-parameter scaling law fitted on a small timestamped reference set, playing the role of the proton–proton reference. The proposal succeeds only if this correction beats the obvious alternative of spending the same small reference set in simpler ways.
Core Claim
Our first claim is diagnostic. Weakly supervised localizers should show a thickness-dependent crossover. In short videos with little background, they under-cover ($R<1$, peak picking). In long, background-heavy videos, they over-cover ($R>1$, context spill). A single global threshold or pooling fraction cannot remove a crossover of this kind; it can only move it.
Our second claim is corrective. Most weakly supervised pipelines pool the top $k$ snippets with $k$ a fixed fraction of video length, and they use outer-inner contrast margins that are a fixed fraction of proposal length. Both choices build in the "transparent" assumption that foreground scales linearly with length. Replacing them with a fitted law should flatten the crossover and raise accuracy at strict overlap thresholds, where boundary errors dominate.
The diagnostic is defined as follows. For a video with true foreground duration $T$ and predicted foreground duration $\hat T$:
$$R_{\text{cov}} = \frac{\hat T}{T} = \underbrace{\frac{\hat n}{n}}_{R_{\text{count}}} \cdot \underbrace{\frac{\bar{\hat s}}{\bar s}}_{R_{\text{span}}},$$
Here $n$ and $\hat n$ are the true and predicted instance counts, and $\bar s$ and $\bar{\hat s}$ are the mean true and predicted instance spans. The split between count and span follows the logic of A dependence of the inclusive production of hadrons with high transverse momenta. That experiment found enhancement for single hadrons but none for back-to-back pairs ($\alpha\approx1.0$). This showed that nuclear matter redistributes momentum rather than creating extra yield. In our setting, a fragmenting localizer can have $R_{\text{cov}}\approx1$ while $R_{\text{count}}>1$ and $R_{\text{span}}<1$. The decomposition catches this case; mAP does not.
The correction replaces the fixed pooling size with
$$k(L) = c\,L^{\alpha},$$
where $L$ is the video length in snippets, and $c$ and $\alpha$ are fitted by least squares on log foreground length against log $L$ over a reference set of about 5% of training videos with timestamps. The same law sets the outer-contrast margin. When $\alpha=1$, this reduces exactly to current practice, so the fitted exponent is itself a test of the linear assumption.
Research Program
The physics papers play clearly separated roles. The centrality-binned $R_{pA}$ measurements in Nuclear-modification factor of charged hadrons at forward and backward rapidity in p+Al and p+Au collisions motivate the diagnostic design. Suppression and enhancement coexist in different regions and grow with nuclear thickness. They became visible only because the ratio was reported by region and centrality rather than integrated. Strong nuclear enhancement in intermediate mass Drell–Yan production models medium effects as corrections that grow with effective thickness ($A^{1/3}$). We take from it only the modelling habit of making the correction an explicit function of thickness. We do not assume any particular exponent carries over to video. That is exactly what $\alpha$ is fitted to find out.
On the vision side, the pieces we modify come from existing methods. AutoLoc introduced the outer-inner contrastive loss. This loss rewards proposals whose inside is more active than a flanking margin, but the margin is a fixed proportion of proposal length. STPN adds a sparsity penalty on attention, which pushes toward narrow peaks. The pairing of STPN-style sparsity with AutoLoc-style contrast is precisely the tug-of-war between the two error modes that our diagnostic is meant to expose. P-MIL contrasts proposals with their surroundings and scores completeness. ACSNet separates action from context with a second branch. Both attack over-coverage architecturally, but neither measures coverage as a function of video length. The duration and shortcut biases reported by Satar et al. and Hira et al. motivate why length should be the stratifying variable.
The work proceeds in three stages. In the first stage, we reproduce four localizers spanning the design space on identical I3D features: CAS thresholding, an STPN-style sparse-attention model, an AutoLoc-style boundary model, and CASE, a clustering-based method. For each, we compute $R_{\text{cov}}$, $R_{\text{count}}$ and $R_{\text{span}}$ in quintile bins of video length and background fraction on THUMOS14 and ActivityNet v1.3.
In the second stage, we fit $(c,\alpha)$ on the reference subset. We insert $k(L)$ into top-$k$ pooling and the fitted margin into the contrast loss of the strongest reproducible base model. We then retrain under the base model's original schedule.
In the third stage, we test transfer. We fit $\alpha$ on THUMOS14 and apply it to ActivityNet, and the reverse. The question is whether the exponent is a property of weak supervision or merely of one dataset.
Baselines
The naive reference point is thresholding the class activation sequence, the per-snippet class score curve, with no boundary modelling. AutoLoc reports this at 13.7% mAP at overlap 0.5 on THUMOS14 and 7.4% on ActivityNet v1.2. We rerun it on our features to anchor the $R$ curves; we expect strong over-coverage in long videos. Among mid-strength methods, TSCN, which uses RGB–flow consensus pseudo-labels, reports 28.7% mAP at 0.5 on THUMOS14. P-MIL reports 47.0% average mAP over overlaps 0.1:0.7 on THUMOS14 and 25.5% average mAP on ActivityNet 1.3.
The strongest existing method in the file is CASE, which reports 57.1% average mAP on THUMOS14 with 2.13M parameters. No reported ActivityNet result for CASE was found on Lacuna. CASE is the right competitor for two reasons. It is the best reported number, and its clustering already tries to separate foreground from background without labels, so it is the method most likely to have little residual coverage bias. We run the official design on the same features and splits, with the same number of epochs and the same hyperparameter search budget for both the base and corrected versions. We count the proposal as beating CASE if the corrected version improves average mAP and mAP at overlap 0.7 over our own CASE reproduction, with the gap exceeding the seed-to-seed spread over five seeds.
The key ablation isolates the scaling exponent. We train the same base model with $\alpha$ fixed at 1 and $c$ re-tuned on the identical reference set. Any gain over this ablation is attributable to length-dependent scaling rather than to access to the reference timestamps.
Why not just…?
Why not just tune a single fixed pooling fraction or threshold on the same reference videos? This is the most serious threat, and it is exactly the $\alpha=1$ ablation. If the stage-one diagnostics show that errors do not change sign with thickness, a single tuned constant should match our method, and we would report that the linear assumption is adequate. The proposal is only worth pursuing if the $R$ curves cross 1. A fixed constant can shift such a crossing but cannot flatten it.
Why not just use the reference timestamps as direct supervision? Twenty-odd annotated videos could instead train the localizer through a segment-level loss, which is ordinary semi-supervision. This may well beat a two-parameter fit in-domain, so we run it as a baseline with the same reference set. Our argument for the scaling law rests on transfer and cost. Two parameters fitted on one dataset can be tested on another without new annotation, and the stage-three cross-dataset experiment decides whether that advantage is real.
Why not just add a head that predicts each video's foreground duration? A learned per-video coverage regressor generalizes our law. But without timestamps it can only learn from video-level labels, which carry no duration signal. With timestamps it reduces to the semi-supervised baseline above. We include it, trained on the reference set, as a higher-capacity competitor. If it wins, the conclusion is that coverage should be predicted rather than scaled, and the diagnostic contribution stands on its own.
Why not just rely on sparsity or two-stream consensus? The file's own evidence argues against these as fixes for coverage. STPN's authors report persistent contextual false positives despite sparsity, and sparsity worsens peak picking. TSCN's consensus still gives imprecise boundaries at high overlap. Our diagnostic makes this concrete, because both methods appear among the reproduced models and should show the predicted opposite-signed errors.
Proposed Experiments
The first decisive check is whether the crossover exists. We will report $R_{\text{span}}$ across length quintiles for each reproduced model on THUMOS14 and ActivityNet v1.3. The corrective half of the proposal is dropped if every model's curve stays on one side of 1, or if the curves are flat within bootstrap intervals over videos. In that case, only the count/span decomposition is published as an evaluation tool.
The second check is whether the fitted exponent departs from 1. The 95% bootstrap interval for $\alpha$ on the reference set must exclude 1. Otherwise, the fitted law is the same as current practice.
The third check is accuracy. The corrected base model must beat both our CASE reproduction and the $\alpha=1$ ablation on average mAP and on mAP at 0.7. Its $R_{\text{span}}$ slope across length bins must also be smaller in magnitude.
The fourth check is transfer. An exponent fitted on THUMOS14 should retain at least half of its in-domain gain on ActivityNet over the fixed-fraction ablation, and the reverse should hold as well. If it does not, the law is dataset-specific, and the semi-supervised baseline is the better use of the reference annotations.
References
- A dependence of the inclusive production of hadrons with high transverse momenta
- Nuclear-modification factor of charged hadrons at forward and backward rapidity in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>p</mml:mi></mml:math> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>+</mml:mo></mml:math> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>Al</mml:mi></mml:math> and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>p</mml:mi></mml:math> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>+</mml:mo></mml:math> <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>Au</mml:mi></mml:math> collisions at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msqrt><mml:msub><mml:mi>s</mml:mi><mml:mrow><mml:mi>N</mml:mi><mml:mi>N</mml:mi></mml:mrow></mml:msub></mml:msqrt><mml:mo>=</mml:mo><mml:mn>200</mml:mn><mml:mspace width="0.16em"/><mml:mi>GeV</mml:mi></mml:mrow></mml:math>
- Strong nuclear enhancement in intermediate mass Drell–Yan production
- AutoLoc: Weakly-Supervised Temporal Action Localization in Untrimmed Videos
- Weakly Supervised Action Localization by Sparse Temporal Pooling Network
- Proposal-Based Multiple Instance Learning for Weakly-Supervised Temporal Action Localization
- ACSNet: Action-Context Separation Network for Weakly Supervised Temporal Action Localization
- Towards Debiasing Frame Length Bias in Text-Video Retrieval via Causal Intervention
- Delta Sampling R-BERT for Limited Data and Low-Light Action Recognition
- Revisiting Foreground and Background Separation in Weakly-supervised Temporal Action Localization: A Clustering-based Approach
- Two-Stream Consensus Network for Weakly-Supervised Temporal Action Localization