Card: focus on the dataset at a glance, metrics and results; link the blog post for the rest

README.md

@@ -24,42 +24,11 @@ configs:
 
 # Perceptron EgoHome: A new benchmark for dense, sub-second egocentric understanding
 
-*The first egocentric benchmark to feature four types of annotations per video, over 2 hours of annotated footage, and segments as short as 100 ms*
+*The first egocentric benchmark to feature four types of annotations per video and segments as short as 100 ms*
 
-This card follows the [EgoHome blog post](https://www.perceptron.inc/blog/egohome), which also has the full-clip videos for each annotation type.
+EgoHome-2h is 137 human-annotated egocentric clips of household chores (1.99 h, 1920×1080, 30 fps), each with hand boxes, subtasks, hand actions and hand captions. The [blog post](https://www.perceptron.inc/blog/egohome) has the full write-up and videos; the scoring code and prompts are on [GitHub](https://github.com/perceptron-ai-inc/egohome_public_release).
 
-As part of our announcement of [Perceptron Egocentric 1.0](https://www.perceptron.inc/blog/introducing-perceptron-egocentric-1-0), we are releasing and open-sourcing EgoHome, a new benchmark containing 2 hours of human-annotated egocentric video of people performing various household tasks.
-
-![The four annotation types on the same 5 s of one clip: hand boxes, subtasks, hand actions and hand captions](assets/egohome_header_2x2.gif)
-
-**Figure 1.** All four of our annotation types on the same five seconds of one clip.
-
-The annotations span four different categories:
-
-- Hand location bounding boxes
-- Subtask annotations
-- Left and Right hand actions (from Ego4D’s closed-vocab schema)
-- Left and Right hand captions.
-
-We release the raw annotations as well as code for evaluation metrics, which we believe better capture the inherent subjectivity of open vocabulary video description.
-
-### A new standard for Egocentric annotations
-
-Rich egocentric annotation is of increasing interest in the physical AI and robotics world, largely because of recent work demonstrating that it is highly effective as a scalable pretraining objective for world action models and vision language action models for [robotic manipulation tasks](https://www.dyna.co/dyna-2) [[1](#references)]. More broadly, recent work has established both the volume and the diversity of pretraining data as essential for reliable control: Physical Intelligence has shown that [compositional generalization](https://www.pi.website/blog/pi07) can emerge from diverse multimodal data and conditioning [[2](#references)], Generalist that [task-specific mastery](https://generalistai.com/blog/gen-1) can be achieved by scaling high-fidelity physical-interaction data [[3](#references)], and Dyna that [scaling egocentric data](https://www.dyna.co/dyna-2) is one viable path to human-level control [[1](#references)].
-
-Despite this increasing attention on egocentric data, it is still not perfectly well-understood which set of annotations serve as the best pretraining objectives or how best to measure the quality of synthetic labeling harnesses for these tasks.
-
-The most widely adopted egocentric benchmark for subtask segmentation, [WGO-Bench](https://macrodata.co/blog/annotating-robot-video-subtasks) [[4](#references)], has been a remarkably useful contribution to the ecosystem and has since become a de facto standard for evaluating the quality of a model or harness on describing the events in an egocentric video. Despite this, it still has several gaps in coverage which we seek to address with this release.
-
-**Perceptron’s EgoHome** benchmark addresses a few gaps left behind by the current state of egocentric benchmarking:
-
-- We introduce benchmarking for four types of annotation all on the same video sources: hand box location, subtask timelines, hand action classification (from the Ego4D closed vocab), and hand action freeform captions.
-- We provide a large number of annotations at the sub-second scale: 3,371 segments under 1 s and 773 under 0.5 s.
-- Finally, we introduce a new flexible metric we call permissible subsegment Jaccard: an LLM-as-judge based technique which accommodates the fact that there are many permissible ways to break down the granularity of a task.
-
-## Benchmark overview
-
-Our benchmark is the first of its kind to combine all four annotation types on every video. This has revealed the spikiness of model capabilities at the frontier, with subtask timeline generation being much stronger than individual hand description for the very best models. We also offer very temporally fine-grained annotations allowing us to analyze how model capabilities break down on much faster action sequences.
+## At a glance
 
 |   | EgoHome (Perceptron) | WGO-Bench |
 |---|---|---|
@@ -72,32 +41,19 @@ Our benchmark is the first of its kind to combine all four annotation types on e
 | Left/right hand captions? | Yes: 5,277 segments | No |
 | Shortest segment | 0.10 s | 0.19 s |
 | Segments < 0.5 s | 773 | 15 |
-| Mean segment length | 2.62 s (subtask 3.57 s) | 5.77 s |
+| Mean subtask length | 3.57 s | 5.77 s |
+| Mean segment length, all types | 2.62 s | 5.77 s (subtasks only) |
 | Flexible time boundary metric? | Yes: permissible subsegment Jaccard | No: one-to-one matching at IoU ≥ 0.75 |
 
 **Table 1.** EgoHome and WGO-Bench at a glance.
 
-### Four annotation types
-
-#### Subtask annotations
+![Cumulative distribution of segment durations on a log scale: EgoHome subtasks, hand actions and hand captions against WGO-Bench](assets/segment_durations.png)
 
-Similar to WGO-Bench, we have open vocab descriptions of the high level goal of what task the subject is performing e.g. “scrubbing the stovetop”. The average video has 14.0 segments.
+**Figure 1.** Cumulative distribution of segment durations (log scale). EgoHome’s hand actions and captions are much shorter than WGO-Bench’s subtasks: about a third of them last under one second.
 
-#### Hand box locations
+## Evaluation metrics
 
-We annotate every hand box with human annotators at 10 fps then apply proprietary smoothing algorithms to extrapolate up to 30 fps.
-
-#### Hand actions
-
-We have annotators choose an accurate hand action label from the closed vocabulary [Ego4D action schema](https://ego4d-data.org/docs/benchmarks/hands-and-objects/) [[5](#references)]; there are 118 total actions to choose from (47 occur in the release).
-
-#### Hand captions
-
-We have annotators write free-form hand captions describing what the hands are doing during the video.
-
-### Evaluation metrics
-
-#### Hand box metric: centroid hit fraction
+### Hand box metric: centroid hit fraction
 
 To measure hand box prediction accuracy, we simply ask each model to produce timestamped bounding boxes for the hand locations throughout the video. For every frame with a human-annotated hand, we check whether the centroid of the model’s box for that hand falls inside the human box, then divide by the number of annotated hand-frames, plus any boxes the model predicts outside the annotated range.
 
@@ -107,7 +63,7 @@ Models that sample fewer frames per second are still scored on every annotated f
 
 **Figure 2.** We stack two seconds of frames in time and check every predicted centroid against the human box.
 
-#### Subtask, action, and hand caption metric: permissible subsegment Jaccard
+### Subtask, action, and hand caption metric: permissible subsegment Jaccard
 
 We developed our subtask correctness metric while attempting to find a fair way to calibrate our annotators. We initially used a metric similar to WGO-Bench’s semantic F1, which performs bipartite matching between the target and prediction and then calculating F1 between the predicted and target sets.
 
@@ -119,7 +75,7 @@ Rather than encouraging models to overfit to one particular choice of temporal g
 
 We call our metric **permissible subsegment Jaccard.**
 
-#### Worked example: the stairs
+### Worked example: the stairs
 
 Let’s say the ground truth labeled the time span 0 to 5s as one segment, “sweeping the stairs”.
 
@@ -127,30 +83,52 @@ The table below demonstrates how various predictions would be graded under our m
 
 | Model output | Permissible subsegment Jaccard | WGO-Bench semantic F1 |
 |---|---|---|
-| A. “sweeping the stairs” 0–5 s | (5 + 5) / (5 + 5) = **1.0** | IoU 1.0, labels match: 1 match → **1.0** |
-| B. “sweeping the top stair” 0–2 s, “sweeping the bottom stair” 2–5 s | both pieces agree: (2 + 3 + 5) / (5 + 5) = **1.0** | IoUs 0.4 and 0.6, both below 0.75: 0 matches of 2 predicted, 1 gold → **0.0** |
-| C. “sweeping the stairs” 0–3 s, “putting the broom away” 3–5 s | first piece agrees: (3 + 3) / (5 + 5) = **0.6** | IoU 0.6, below 0.75 → **0.0** |
-| D. “carrying the broom” 0–5 s | judge says it’s a different action → **0.0** | timing matches but the label doesn’t → **0.0** |
+| A. 0–5 s: “sweeping the stairs” | (5 + 5) / (5 + 5) = **1.0** | IoU 1.0, labels match: 1 match → **1.0** |
+| B. 0–2 s: “sweeping the top stair”; 2–5 s: “sweeping the bottom stair” | both pieces agree: (2 + 3 + 5) / (5 + 5) = **1.0** | IoUs 0.4 and 0.6, both below 0.75: 0 matches of 2 predicted, 1 gold → **0.0** |
+| C. 0–3 s: “sweeping the stairs”; 3–5 s: “putting the broom away” | first piece agrees: (3 + 3) / (5 + 5) = **0.6** | IoU 0.6, below 0.75 → **0.0** |
+| D. 0–5 s: “carrying the broom” | judge says it’s a different action → **0.0** | timing matches but the label doesn’t → **0.0** |
 
-**Table 2.** Four model outputs for one human segment, “sweeping the stairs” from 0 to 5 s, scored by permissible subsegment Jaccard and WGO-Bench semantic F1. The ground-truth track is the human label; the model track is the prediction. Solid segments agree with the ground truth under the judge; dashed segments do not.
+**Table 2.** Four cases of model output for one human-labeled (ground truth) segment, “sweeping the stairs” from 0 to 5 s, scored by permissible subsegment Jaccard and WGO-Bench semantic F1. The ground-truth track is the human label; the model track is the prediction. Solid segments agree with the ground truth under the judge; dashed segments do not.
 
 While calibrating annotators, we frequently ran into Case B, as illustrated in Table 2: a valid, more detailed breakdown gets a score of zero under one-to-one matching and full credit under ours. Case C shows we still give partial credit for the part the model got right, where a fixed IoU cutoff gives nothing. (WGO-Bench’s semantic F1 pairs segments one-to-one, counts a pair only when its temporal IoU is at least 0.75 and the judge accepts the labels).
 
 We found that this metric adhered much more closely to subjective measures of fairness when evaluating human annotators’ performance during calibration, thus we opted to release it as our official metric for the benchmark.
 
-The formal definitions of both metrics are in the [appendix](#metric-formulas), and the scoring code is in our [GitHub repository](https://github.com/perceptron-ai-inc/egohome_public_release).
+The formal definitions follow below, and the scoring code is in our [GitHub repository](https://github.com/perceptron-ai-inc/egohome_public_release).
+
+### Metric formulas
+
+#### Centroid hit fraction
+
+Mathematically this can be expressed as
+
+$$
+\mathrm{CHF} = \frac{\displaystyle\sum_{(f,h) \in G} \mathbb{1}\left[\mathrm{Centroid}(\hat{p}_{f,h}) \in g_{f,h}\right]}{|G| + |P_{\text{out}}|} \tag{1}
+$$
+
+Where $G$ is the set of human-annotated frames at 30 fps and $g_{f,h}$ is the ground truth box for hand $h$ on frame $f$. $\hat{p}_{f,h}$ is the model’s box for hand $h$ at frame $f$: either a box directly emitted for that frame, or the linear interpolation between its two nearest boxes; frames outside the span the model covered count as misses. $P_{\text{out}}$ is the set of frames for which the model has a box but the ground truth doesn’t. The indicator in the numerator is 1 when the centroid of $\hat{p}_{f,h}$ falls inside $g_{f,h}$.
+
+#### Permissible subsegment Jaccard
+
+It can be expressed mathematically as follows:
+
+$$
+\mathrm{PSJ}(P, T) = \frac{\displaystyle\sum_{p \in P} m(p, T) + \sum_{t \in T} m(t, P)}{\displaystyle\sum_{p \in P} |I(p)| + \sum_{t \in T} |I(t)|} \tag{2}
+$$
 
-## How different models stack up
+$$
+m(s, S) = \Bigl|\, I(s) \cap \bigcup_{\substack{s' \in S \\ J(s, s') = 1}} I(s') \,\Bigr| \tag{3}
+$$
 
-We compare Perceptron Egocentric, at its default 4 fps and an 8 fps option, with GPT-6 Astra, Gemini 3.8 Flash and Elorian’s Glimpse, each at 4 fps and 1 fps, and with Pantheon’s Argus and NVIDIA’s FLORA. Subtasks, hand actions and hand captions are scored with permissible subsegment Jaccard, and hand boxes with centroid hit fraction; higher is better for both.
+Let $P$ and $T$ be the predicted and ground-truth segments. Each segment $s$ has a time interval $I(s)$ and a label, and $|\cdot|$ is duration. $J(s, s') \in \{0, 1\}$ is the LLM judge’s verdict on whether two labels are permissible matches: the same action, or one a subtask of the other. $m(s, S)$ is the portion of $s$’s duration that overlaps some segment in $S$ that the judge accepts.
 
-The Perceptron Egocentric harness matches or beats every other system on subtask labeling and is close to GPT-6 Astra on hand boxes, but trails it on hand actions and captions. Generally speaking, most systems struggle with identifying individual hand activities, with the 3 hand-related portions of this benchmark quite far from being saturated.
+## Results
 
 ![Subtask score against subtask cost per video-hour, log scale; the dashed line is the Pareto frontier](assets/subtask_pareto.png)
 
 **Figure 3.** Both Perceptron Egocentric settings sit on the Pareto frontier; nothing we measured beats them on both score and cost.
 
-### EgoHome-2h (137 clips)
+### EgoHome results
 
 | System | Subtask ↑ | Hand action ↑ | Hand caption ↑ | Hand box ↑ | Subtask $ / video-hour ↓ |
 |---|---:|---:|---:|---:|---:|
@@ -170,8 +148,6 @@ The Perceptron Egocentric harness matches or beats every other system on subtask
 
 ### Fine-grained results by gold segment length
 
-Our benchmark is the only one with sufficient scale of annotations necessary to offer performance breakdowns at varying timescales. This reveals that specialist systems like NVIDIA's FLORA hold up much better at shorter timescales than some of the other models which fare well on the overall metric.
-
 **Subtask**
 
 | System | All ↑ | < 0.5 s ↑ | 0.5–1 s ↑ | > 1 s ↑ |
@@ -223,33 +199,6 @@ Our benchmark is the only one with sufficient scale of annotations necessary to
 
 **Table 6.** Hand caption score by the length of the gold segment. Best in bold.
 
-## How we built EgoHome
-
-### Annotation methodology
-
-We begin by synthetically labeling videos then having humans edit the initial annotations. Next, we had a member of our research team review every clip by hand. Due to the quality of our initial annotations and our proprietary review pipeline, we were able to get the overall average time to re-annotate a video down to 116 minutes per minute of video across all four annotation types.
-
-| Annotation type | Active minutes per minute of video | Agreement of the synthetic seed with the final human labels |
-|---|---|---|
-| Hand boxes | 41.7 | 0.932 (centroid hit fraction) |
-| Subtasks | 17.1 | 0.766 (permissible subsegment Jaccard) |
-| Hand actions | 24.6 | 0.613 (permissible subsegment Jaccard) |
-| Hand captions | 32.6 | 0.688 (permissible subsegment Jaccard) |
-| All four | 115.9 |  |
-
-**Table 7.** Annotation effort per minute of video, and how closely our synthetic seed labels agreed with the final human labels.
-
-#### Ensuring annotator quality
-
-Every annotator undergoes a calibration where they have to achieve high agreement with a clip labeled by a member of our research staff. They are then allowed to re-attempt the calibration clip up to 3 times, with feedback after each failed attempt. They are allowed to see what parts of their answers were incorrect, but they are not allowed to see the ground truth.
-
-After their initial calibration, crowdsourced annotators then go through a 10-clip calibration against each other, where they are graded by how high their interannotator agreement is compared to their agreement with the original labels. Via this process we chose 5 annotators (from a pool of 50) to serve as the annotators on our evaluation set, keeping lower performing members of our annotation pool to work on other tasks.
-
-## Open-source release
-
-- 🤗 [Hugging Face: EgoHome dataset](https://huggingface.co/datasets/PerceptronAI/EgoHome)
-- 🐙 [GitHub: metrics and judge prompts](https://github.com/perceptron-ai-inc/egohome_public_release)
-
 ## Using the dataset
 
 ```python
@@ -278,196 +227,18 @@ There is one row per clip. Times are seconds from the clip start. Boxes are `[x_
 | `hand_actions` | `{left: [{start, end, label}], right: [...]}`; labels are Ego4D FHO verb classes |
 | `hand_captions` | `{left: [{start, end, label}], right: [...]}` |
 
-## Appendix
-
-### Metric formulas
-
-#### Centroid hit fraction
-
-Mathematically this can be expressed as
-
-$$
-\mathrm{CHF} = \frac{\displaystyle\sum_{(f,h) \in G} \mathbb{1}\left[\mathrm{Centroid}(\hat{p}_{f,h}) \in g_{f,h}\right]}{|G| + |P_{\text{out}}|}
-$$
-
-Where G is the set of (frame, hand) pairs with a human box at 30 fps and g(f,h) is that box. p̂(f,h) is the model’s box for hand h at frame f: its own box at that frame, or the linear interpolation between its two nearest boxes; frames outside the span the model covered count as misses. P_out is the set of predicted boxes outside the frames the human labels cover, and the indicator is 1 when the centroid of p̂(f,h) falls inside g(f,h).
-
-#### Permissible subsegment Jaccard
-
-It can be expressed mathematically as follows:
-
-$$
-\mathrm{PSJ}(P, T) = \frac{\displaystyle\sum_{p \in P} m(p, T) + \sum_{t \in T} m(t, P)}{\displaystyle\sum_{p \in P} |I(p)| + \sum_{t \in T} |I(t)|}
-$$
-
-$$
-m(s, S) = \Bigl|\, I(s) \cap \bigcup_{\substack{s' \in S \\ J(s, s') = 1}} I(s') \,\Bigr|
-$$
-
-Let P and T be the predicted and ground-truth segments. Each segment s has a time interval I(s) and a label, and |·| is duration. J(s,s′) ∈ {0,1} is the LLM judge’s verdict on whether two labels are permissible matches: the same action, or one a subtask of the other. m(s,S) is the portion of s’s duration that overlaps some segment in S that the judge accepts.
-
-### Judge prompt
-
-For subtasks and hand captions, an LLM judge (Gemini 3.5 Flash) decides whether two overlapping labels describe the same activity; hand actions are matched on their Ego4D verb. This is the judge’s prompt for subtasks. For hand captions, “subtask descriptions” reads “hand activity captions”. The prompts each model receives are in our [GitHub repository](https://github.com/perceptron-ai-inc/egohome_public_release).
-
-```text
-You are grading a human annotator against a gold reference for egocentric video subtask descriptions. The two labels describe time segments that overlap; decide if they are talking about the SAME underlying activity.
-Gold label: "<gold label>"
-Annotator label: "<predicted label>"
-Be lenient about WORDING, OBJECTS, and SURFACES: match=true across synonyms, paraphrases, different objects/tools/surfaces, and finer-vs-coarser detail, PROVIDED the same physical hand ACTION is being performed (e.g. "vacuuming the stairs with the hose" ~ "vacuuming the floor" — both vacuuming; "wash the dishes" ~ "scrub the plate" — both washing/scrubbing).
-But the ACTION itself must be the same: return match=false when the manipulation differs, EVEN IF the same object is mentioned (e.g. "vacuuming the stairs with the hose" vs "holding the vacuum cleaner" — vacuuming vs holding; "wiping a window" vs "opening a drawer"), or when one is empty/placeholder.
-A label that names no specific physical action never matches a specific one: generic descriptions such as "doing chores", "cleaning up", "working in the kitchen", "doing the task" or "using the hands" are match=false against any specific activity.
-When genuinely ambiguous, lean match=false.
-Return only JSON: {"match": true|false, "rationale": "short reason"}
-```
-
-### Degenerate baselines
-
-In order to ensure the metric isn’t trivially gameable, we chose a few degenerate baselines and determined how they'd be rated by the metric. We chose three types of generic baseline:
-
-1. Choosing a generic label like "doing chores" and applying it to the whole timeline
-2. Choosing a generic label and applying at fixed 2.5s windows
-3. Taking the most common/longest true clip from the ground truth and filling the whole video timeline with it as a single task.
-
-| Answer (no video used) | Subtask | Hand action | Hand caption |
-|---|---:|---:|---:|
-| One segment for the whole clip, generic label | 0.000 | 0.175 | 0.000 |
-| Fixed 2.5 s windows, generic label | 0.000 | 0.175 | 0.000 |
-| One segment for the whole clip, the clip’s own most common true label | 0.528 | 0.449 | 0.499 |
-
-**Table 8.** Scores of three degenerate answers that use no video.
-
-Generic labels earn nothing on subtasks and captions. A constant “hold” earns 0.175 on hand actions, the share of time hands really are holding. These degenerate baselines represent a floor of how good a trivial baseline is at the task. The table below shows which models exceed the trivial baseline on the three tasks.
-
-| System | Subtask | Hand action | Hand caption |
-|---|---:|---:|---:|
-| Perceptron Egocentric, 4 fps (default) | 0.787 ✓ | 0.421 ✗ | 0.613 ✓ |
-| Perceptron Egocentric, 8 fps (option) | 0.798 ✓ | 0.421 ✗ | 0.626 ✓ |
-| GPT-6 Astra, 4 fps frames | 0.797 ✓ | 0.498 ✓ | 0.749 ✓ |
-| GPT-6 Astra, 1 fps frames | 0.775 ✓ | 0.479 ✓ | 0.724 ✓ |
-| Gemini 3.8 Flash, 4 fps video | 0.663 ✓ | 0.405 ✗ | 0.558 ✓ |
-| Gemini 3.8 Flash, 1 fps video | 0.639 ✓ | 0.390 ✗ | 0.557 ✓ |
-| Glimpse (Elorian), 4 fps video | 0.689 ✓ | 0.390 ✗ | 0.560 ✓ |
-| Glimpse (Elorian), 1 fps video | 0.590 ✓ | 0.353 ✗ | 0.560 ✓ |
-| Argus (Pantheon), GPT-6 Astra | 0.367 ✗ | 0.336 ✗ | 0.593 ✓ |
-| Argus (Pantheon), Midcentury's settings: GPT-6.1 Sol | 0.315 ✗ | 0.327 ✗ | 0.590 ✓ |
-| FLORA (NVIDIA) | 0.747 ✓ | — | — |
-| *Best trivial baseline* | 0.528 | 0.449 | 0.499 |
-
-**Table 9.** Which systems exceed the best trivial baseline on each task. ✓ exceeds the best trivial baseline for that task; ✗ does not. n/r = not reported for that setting. — = the system does not produce that annotation.
-
-### Annotation statistics
-
-#### Subtasks
-
-The distribution of durations is below:
-
-| Subtask segments | EgoHome |
-|---|---:|
-| Segments | 1,924 |
-| Shortest | 0.16 s |
-| Median | 2.50 s |
-| Mean | 3.57 s |
-| 90th percentile | 7.34 s |
-| Longest | 55.1 s |
-
-**Table 10.** Subtask segment statistics.
-
-| Duration | EgoHome |
-|---|---:|
-| < 0.5 s | 30 (1.6%) |
-| 0.5–1 s | 120 |
-| 1–2 s | 564 |
-| 2–5 s | 829 |
-| 5–10 s | 277 |
-| > 10 s | 104 |
-
-**Table 11.** Subtask segments by duration.
-
-#### Hand boxes
-
-Here is the percentage of time the right and left hand are on screen in the corpus:
-
-| Hand on screen | % of frames |
-|---|---:|
-| Left | 96.7% |
-| Right | 96.6% |
-| Both | 94.3% |
-| Either | 99.0% |
-| Neither | 1.0% |
-| Boxes | 417,912 on 216,225 frames |
-
-**Table 12.** Share of frames with each hand on screen.
-
-#### Hand actions
-
-| Hand actions | Left | Right |
-|---|---|---|
-| Segments per clip | 17.3 | 19.4 |
-| Median duration | 1.45 s | 1.34 s |
-| Mean duration | 2.49 s | 2.39 s |
-| < 0.5 s | 171 | 248 |
-| < 1 s | 731 | 939 |
-| Distinct labels (both hands) | 47 |  |
-| Video where left and right have the same action | 42.7% |  |
-
-**Table 13.** Hand action segments by hand.
-
-| Most frequent verbs | Segments |
-|---|---:|
-| hold | 802 |
-| take | 615 |
-| put | 595 |
-| fold | 427 |
-| move | 310 |
-| arrange | 257 |
-| adjust | 227 |
-| smooth | 214 |
-| lift | 207 |
-| wash | 163 |
-
-**Table 14.** The ten most frequent hand action verbs.
-
-#### Hand captions
-
-| Hand captions | Left | Right |
-|---|---|---|
-| Segments per clip | 18.4 | 20.1 |
-| Median duration | 1.55 s | 1.49 s |
-| Mean duration | 2.50 s | 2.38 s |
-| < 0.5 s | 145 | 179 |
-| < 1 s | 707 | 844 |
-| Distinct labels (both hands) | 3,948 |  |
-| Video where left and right have the same caption | 29.3% |  |
+## License
 
-**Table 15.** Hand caption segments by hand.
+The EgoHome dataset is released under the Creative Commons Attribution-NonCommercial 4.0 license (CC BY-NC 4.0). As an additional permission, anyone, including commercial organizations, may use EgoHome to evaluate models and publish the results. Other commercial uses, such as selling the data or training commercial models on it, are not permitted.
 
 ## Citation
 
 ```bibtex
-@misc{perceptron2026egohome,
-  title={Perceptron {EgoHome}: A new benchmark for dense, sub-second egocentric understanding},
-  author={{PerceptronAI}},
-  year={2026},
-  month={nov},
-  url={https://www.perceptron.inc/blog/egohome}
+@article{perceptron2026egohome,
+  author = {PerceptronAI},
+  title = {Perceptron EgoHome: A new benchmark for dense, sub-second egocentric understanding},
+  journal = {PerceptronAI Research Blog},
+  year = {2026},
+  note = {https://perceptron.inc/blog/egohome},
 }
 ```
-## License
-
-The EgoHome dataset is released under the Creative Commons Attribution-NonCommercial 4.0 license (CC BY-NC 4.0). As an additional permission, anyone, including commercial organizations, may use EgoHome to evaluate models and publish the results. Other commercial uses, such as selling the data or training commercial models on it, are not permitted.
-## Acknowledgements
-
-The raw footage for this data was collected by [HumanSignal](https://humansignal.com/), a human data provider specializing in high-quality data collection and annotation for AI teams.
-
-## References
-
-1. Dyna Robotics. Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models. August 2026. [dyna.co/dyna-2](https://www.dyna.co/dyna-2)
-
-2. Physical Intelligence. A Steerable Model with Emergent Capabilities. April 2026. [pi.website/blog/pi07](https://www.pi.website/blog/pi07)
-
-3. Generalist. GEN-1: Scaling Embodied Foundation Models to Mastery. April 2026. [generalistai.com/blog/gen-1](https://generalistai.com/blog/gen-1)
-
-4. Macrodata Labs. Segmenting Robot Video into Actionable Subtasks. June 2026. [macrodata.co/blog/annotating-robot-video-subtasks](https://macrodata.co/blog/annotating-robot-video-subtasks)
-
-5. K. Grauman, A. Westbury, E. Byrne, et al. Ego4D: Around the World in 3,000 Hours of Egocentric Video. CVPR 2022. [arXiv:2110.07058](https://arxiv.org/abs/2110.07058)