Human Action Classification is live · free · last verified 30 Oct 2026 · viability healthy
Human Action Classification
MC3-18 fine-tuned on UCF-101: an Apache-2.0 PyTorch checkpoint for 101-class trimmed-clip video action recognition at 87.05% top-1.
If your footage is short, single-action and UCF-101-shaped, this is the strongest free checkpoint in its own family: 87.05% top-1 at 11.5M parameters is a lean trade, and the evaluation protocol is documented well enough to reproduce. The scope limits are the real decision. One label per clip, no temporal localisation, and nothing outside UCF-101's fixed 101 classes. If inference speed or footprint matters more than the last few accuracy points, take the project's R3D-18 at 83.43%. If you need untrimmed or open-vocabulary video understanding, look at heavier backbones such as I3D or R(2+1)D in the torchvision video lineup, or a video-language model — neither is what this checkpoint does.
Verdict generatedVerified 7h ago · liveness 73/100 · cite: rightaichoice.com/tools/human-action-classification
- Researchers who need a documented, directly comparable UCF-101 checkpoint
- CV engineers prototyping trimmed-clip action classification on CPU-only hardware
- Educators teaching 3D CNN video pipelines with a small model
- Students and hobbyists who need a free, Apache-2.0-licensed video classifier
- Videos containing multiple actions that need temporal localization
- Open-vocabulary recognition outside UCF-101's fixed 101 classes
- Untrimmed, multi-person or surveillance-style footage — performance is unverified by the author
We scan live Reddit threads, YouTube comments, X posts, G2 reviews and other communities — and hand you an honest verdict in under a minute.
- Honest verdict, not marketing
- Real pros & cons from real users
- Attributed quotes with receipts
3 free scans · no card needed
Skip this checkpoint if your video contains several actions, needs time-stamped localization, or falls outside UCF-101's 101 classes — you need a detection or open-vocabulary model here, not a whole-clip classifier.
The package is not on PyPI, so you pay in setup time: git clone plus pip install -e . from source before a single prediction runs.
The checkpoint itself is $0 under Apache-2.0 and needs no account to download. Your actual spend is compute and any Hub storage you use: public repository storage is free at basic volume, while Hugging Face charges $12/TB/mo for public and $18/TB/mo for private repositories at the base tier, and a Solo PRO account is $9/month. For a 11.5M-parameter state dict that is well under a gigabyte, storage cost is effectively rounding error next to the engineering time.
In short
Human Action Classification — MC3-18 fine-tuned on UCF-101: an Apache-2.0 PyTorch checkpoint for 101-class trimmed-clip video action recognition at 87.05% top-1. Best for Researchers who need a documented, directly comparable UCF-101 checkpoint, CV engineers prototyping trimmed-clip action classification on CPU-only hardware, Educators teaching 3D CNN video pipelines with a small model. Free to use.
What is Human Action Classification?
Human Action Classification — MC3-18 fine-tuned on UCF-101: an Apache-2.0 PyTorch checkpoint for 101-class trimmed-clip video action recognition at 87.05% top-1. Best for Researchers who need a documented, directly comparable UCF-101 checkpoint, CV engineers prototyping trimmed-clip action classification on CPU-only hardware, Educators teaching 3D CNN video pipelines with a small model. Free to use.
MC3-18 for UCF-101 is a mixed 2D/3D convolution video classification checkpoint published by the human-action-classification project on Hugging Face Hub. You pass it a 16-frame clip sampled uniformly across a video, and it returns one action label from UCF-101's 101 classes plus a confidence and a top-k list. It is the best-accuracy entry in the project's UCF-101 lineup, paired with a faster R3D-18 sibling at 83.43%. The published numbers are concrete: 87.05% top-1 on UCF-101 split 1's official 3,783-video test list, macro precision 87.34%, macro recall 86.95%, macro F1 86.18%, roughly 11.5M parameters, best epoch 170 of 200. Those figures come from VideoTrainer.validate() in hac.video.training.train — a single center clip per video, no test-time augmentation or multi-crop averaging. Weights initialise from a Kinetics-400 checkpoint, then train on 9,537 UCF-101 split-1 videos with SGD (momentum 0.9, lr 0.001, weight decay 0.0005, StepLR step_size=20 gamma=0.1). Loading runs through huggingface_hub and PyTorch: fetch config.json and mc318-ufc101-split-1.pth, rebuild Video3DCNN, load the state dict. The repo's VideoPredictor wraps frame sampling, the aspect-preserving resize to roughly 128x171, the 112x112 center crop, Kinetics-400 normalisation and the forward pass, so predict_video returns top_class, top_confidence and a top-k list in a few lines on CPU. Who it suits: computer-vision researchers who want a documented, directly comparable UCF-101 baseline; engineers prototyping trimmed-clip activity tagging on CPU-only hardware; and educators who want a small, readable 3D CNN for teaching video pipelines. It is not open-vocabulary and does not localise actions in time — one clip in, one label out.
Key features
- MC3-18 architecture mixing 2D and 3D convolutions across stages
- Fine-tuned for 101-class UCF-101 action recognition on official split 1
- 87.05% top-1 accuracy on the 3,783-video split-1 test list
- Macro precision 87.34%, macro recall 86.95%, macro F1 86.18%
- Roughly 11.5M parameters; best epoch 170 of 200
- Best accuracy in this project's UCF-101 model zoo
- Pretrained init from a Kinetics-400 checkpoint
- 16 frames sampled uniformly across the full video
- Aspect-preserving resize to ~128x171, center crop to 112x112
- Kinetics-400 normalization statistics in the transform chain
- Clip-level output over 101 action classes with confidence and top-k list
- VideoPredictor wraps frame sampling, transforms and forward pass end-to-end
- predict_video returns top_class, top_confidence and a top-k prediction list
- Runs on CPU via device='cpu' in the published usage snippet
- Apache-2.0 licensed checkpoint and config on Hugging Face Hub
Behind the verdict
Treat this as a research artefact, not a product. What you get is a config.json, a 11.5M-parameter state dict and a model card that tells you exactly how the numbers were produced — which is more than most published video baselines bother to document. The card is unusually honest about provenance: it flags that the augmentation settings and batch size are the training script's CLI defaults, not values independently confirmed for this exact run, and it notes no separate run-config was saved alongside the checkpoint. That kind of disclosure is worth a lot when you are deciding whether to trust a 87.05% figure. The strengths are narrow and real. It runs on CPU via device='cpu' in the published snippet, so you can get a prediction without a GPU. The metrics are traceable: single center clip, 16 uniform frames, 112x112 crop, Kinetics-400 normalisation, no test-time augmentation. And the project ships a second checkpoint, R3D-18 at 83.43%, trained with the same pipeline, so you can compare two backbones on identical footing rather than reading numbers from different papers. Where it disappoints is scope. One label per clip means a video with several actions gets a single guess. Sixteen uniformly sampled frames will miss brief or rapid actions buried in a longer clip. The class list is UCF-101's 101 human actions and nothing else — there is no open-vocabulary path. And the card itself says performance on untrimmed, multi-person or surveillance-style footage is unverified by the author, which is the sentence that should govern your expectations. Operationally, expect friction. The package is not published on PyPI, so you install from source with git clone plus pip install -e . on the dronefreak/human-action-classification repo. There is no hosted endpoint, no SDK, no dashboard — the Hugging Face repo is the distribution channel. If you want a managed inference API, this is the wrong artefact; if you want to read the code, change the transform chain and rerun the benchmark yourself, it is exactly right.
Models under the hood
as of 2026-10-24
Limitations
- Fixed to UCF-101's 101 action classes, evaluated on split 1's official test list of 3,783 videos.
- Metrics come from a single center-cropped 16-frame clip at 112x112 with no test-time augmentation or multi-crop averaging, and uniform 16-frame sampling can miss brief or rapid actions in longer clips.
- It is the best-accuracy model in this project's own UCF-101 lineup (87.05% vs R3D-18 at 83.43%), not a general-purpose or state-of-the-art action recognizer.
- Distributed as an Apache-2.0 PyTorch checkpoint on Hugging Face Hub rather than a hosted service, and the package is not yet published on PyPI, so you install from source.
- The model card itself states that augmentation settings and batch size were not independently confirmed for this exact run — no separate run-config was saved alongside the checkpoint.
as of 2026-10-30
Researching Human Action Classification? Get your full AI stack in 60 seconds.
Free, no signup — tell us your goal and get tools matched to your budget & existing stack.
Is Human Action Classification still active?
Human Action Classification is live · free · last verified 30 Oct 2026 · viability healthy.
Viability score
How well maintained and how widely used is Human Action Classification? Built from what the vendor actually publishes (docs, changelog, tutorials, integrations, pricing), whether the site is live, and how much real users discuss it. How we calculate this
Last calculated: October 2026
How we score →What do users say about Human Action Classification?
What people actually say about Human Action Classification — is it worth it?
Our community sentiment research written Jul 5, 2026 is more than 90 days old, so it is being re-reviewed against current data and is not shown.
How we verified this
Every listed fact about Human Action Classification— pricing, features, links, status — is re-read from the vendor's own pages by our verification pipeline and each change is logged with a date; the last pass ran on 30 Oct 2026. Pages mix model-assisted synthesis with human review; see how we verify.
Verification history
We have re-verified Human Action Classification 10 times since . Each pass re-reads the vendor's own pages and re-checks every listed field against that evidence; passes where nothing had changed are marked as such.
- — changed switching in, hidden costs, summary, our verdict, who should skip it, features, description, switching out, limitations, setup time, pricing fit
- — changed pricing page
- — changed limitations, pricing page, platforms
- — changed limitations, pricing page
- — changed who it suits, description, our verdict, features, hidden costs, limitations, switching in, switching out, who should skip it, pricing tiers, pricing fit, setup time, summary, use cases
- — changed limitations, platforms, pricing page
Showing the 6 most recent of 10 verification passes.
Free to cite with attribution — this page re-verifies continuously.
How much does Human Action Classification cost in 2026?
Human Action Classification: Free to use. Pricing last verified 30 Oct 2026.
12-month cost
Project the real annual outlay, including the implied monthly cost when only an annual tier is published.
Vendor list price only. Add-on usage, seat overages, and contract minimums are surfaced under Hidden costs & gotchas.
Where the pricing makes sense
The company stage and team size where Human Action Classification's pricing actually pencils out — and where peers do it cheaper.
The checkpoint itself is $0 under Apache-2.0 and needs no account to download. Your actual spend is compute and any Hub storage you use: public repository storage is free at basic volume, while Hugging Face charges $12/TB/mo for public and $18/TB/mo for private repositories at the base tier, and a Solo PRO account is $9/month. For a 11.5M-parameter state dict that is well under a gigabyte, storage cost is effectively rounding error next to the engineering time.
Has Human Action Classification's pricing changed?
No plan price for Human Action Classification has moved since we began logging in Aug 2026; the pricing page is re-read on every verification pass.
Who should pick Human Action Classification?
Pick Human Action Classification if you are Researchers who need a documented, directly comparable UCF-101 checkpoint, CV engineers prototyping trimmed-clip action classification on CPU-only hardware, Educators teaching 3D CNN video pipelines with a small model, Students and hobbyists who need a free, Apache-2.0-licensed video classifier.
Look elsewhere if you are Videos containing multiple actions that need temporal localization, Open-vocabulary recognition outside UCF-101's fixed 101 classes, Untrimmed, multi-person or surveillance-style footage — performance is unverified by the author, Teams that need a managed inference endpoint rather than a downloadable checkpoint.
Skip it if Skip this checkpoint if your video contains several actions, needs time-stamped localization, or falls outside UCF-101's 101 classes — you need a detection or open-vocabulary model here, not a whole-clip classifier.
GeneratedReal-world workflow fit
Concrete scenarios for the personas Human Action Classification actually fits — and what changes day-one when you adopt it.
You are comparing a new lightweight backbone against a known UCF-101 baseline and need numbers produced by an identical evaluation loop rather than scraped from disparate papers.
Outcome: You pull config.json and mc318-ufc101-split-1.pth via hf_hub_download, rebuild Video3DCNN, and read the model card's protocol — single center clip, 16 uniform frames, 112x112 — so your reproduction is apples-to-apples with the published 87.05% top-1.
You have a folder of short sports clips on a laptop with no GPU and want a first-pass activity tag for each.
Outcome: You construct VideoPredictor(model_path=weights_path, device='cpu'), call predict_video with top_k=5, and get top_class, top_confidence and a five-item prediction list per clip without provisioning a GPU.
You want students to read a complete 3D CNN pipeline — frame sampling, resize, crop, normalisation, forward pass — without a 100M-parameter codebase in the way.
Outcome: The 11.5M-parameter MC3-18 and the VideoPredictor wrapper give a readable end-to-end example, and the model card's training configuration table (SGD momentum 0.9, lr 0.001, StepLR step_size=20) doubles as lecture material.
Use cases
- Classify human actions in short trimmed video clips for sports analysis
- Establish a documented UCF-101 baseline to compare new action recognition architectures against
- Tag video segments with activity labels in automated annotation pipelines
- Prototype a low-compute action recognition system for CPU-only or embedded hardware
- Teach video classification with a small, readable 3D CNN pipeline
Setup time & first value
How long it actually takes to get something useful out of Human Action Classification — broken out by persona, not the marketing-page minute.
Roughly 15-30 minutes to a first prediction: clone dronefreak/human-action-classification, pip install -e ., then hf_hub_download the config and 11.5M-parameter checkpoint. CPU inference works out of the box with the published snippet. Budget an extra hour if you want to reproduce the 87.05% benchmark, since the exact batch size and augmentation settings were not saved with the checkpoint.
Switching to or from Human Action Classification
How to bring data in from common predecessors and how to get it back out — written for the switcher, not the buyer.
- →From the project's R3D-18 UCF-101 checkpoint: swap the repo_id and weights filename, keep Video3DCNN and VideoPredictor unchanged — same pipeline, 83.43% to 87.05% top-1.
- →From a torchvision mc3_18 Kinetics-400 model: load mc318-ufc101-split-1.pth into Video3DCNN with num_classes=101 instead of 400.
- ↗To R3D-18 in the same zoo: point VideoPredictor at the R3D-18 weights for a lighter, faster baseline at 83.43% top-1.
- ↗To torchvision video models such as I3D or R(2+1)D: keep the 16-frame / 112x112 / Kinetics-400-normalisation transform chain and swap the backbone.
- ↗To a video-language model: needed only if you require open-vocabulary labels or temporal localisation, which this checkpoint does not provide.
Integrations
Resources & guides
- Resourcehuggingface.co
Mc3 18 Ucf101 · Human Action Classification
Helpful link from huggingface.co
- Resourcegithub.com
Human Action Classification · Human Action Classification
Helpful link from github.com
- Documentationhuggingface.co
Index · Human Action Classification
Full product docs from huggingface.co
- Documentationhuggingface.co
Docs · Human Action Classification
Full product docs from huggingface.co
Tutorials & Learning
YouTube returned 6 videos for “Human Action Classification”, and we withheld 6: 6 did not mention Human Action Classification. We are showing none, because we could not prove any of them are about Human Action Classification.
Official links
What are the best Human Action Classification alternatives?
The closest alternatives to Human Action Classification we list are Cerul, Parseq and SAM 3 & 3D. Ranked by direct product-type match, not category overlap.
| Alternative | Pricing | Starts at | Last verified |
|---|---|---|---|
| Cerul | Freemium | Pricing not published | — |
| Parseq | Free | Free | — |
| SAM 3 & 3D | Free | Free | — |
Alternatives to Human Action Classification
View all
Cerul
Cerul turns video archives into semantic video search — ask in plain words, get the exact timestamped moment.
Parseq
Free scene text recognition in the browser with PARSeq's permuted autoregressive OCR model
SAM 3 & 3D
SAM 3 does text-prompt, open-vocabulary segmentation and tracking in images and video — free to try in Meta's playground.
Tools that pair well with Human Action Classification
Common stack mates teams adopt alongside Human Action Classification, with the specific reason each pairing earns its keep.
Cerul
Cerul turns video archives into semantic video search — ask in plain words, get the exact timestamped moment.
Parseq
Free scene text recognition in the browser with PARSeq's permuted autoregressive OCR model
SAM 3 & 3D
SAM 3 does text-prompt, open-vocabulary segmentation and tracking in images and video — free to try in Meta's playground.
Human Action Classification vs its closest rival
Human Action Classification's closest head-to-head is with Landr Mastering — These tools serve entirely different domains: Human Action Classification is a free, open-source model for video action recognition, ideal for researchers and engineers needing a UCF-101 baseline. LANDR Mastering is a freemium AI mastering service for musicians, offering professional-grade audio polish via web, DAW plugin, or mobile app. Your choice depends on whether your task is video understanding or music production—they are not substitutes. Verdict generated
Head-to-head comparisons
Human Action Classification vs Landr Mastering
These tools serve entirely different domains: Human Action Classification is a free, open-source model for video action recognition, ideal for researchers and engineers needing a UCF-101 baseline. LANDR Mastering is a freemium AI mastering service for musicians, offering professional-grade audio polish via web, DAW plugin, or mobile app. Your choice depends on whether your task is video understanding or music production—they are not substitutes.
GeneratedHuman Action Classification vs Splice
These tools are incomparable: Human Action Classification is a free research model for video action recognition, while Splice is a freemium platform for music production. Choose Human Action Classification if you need a PyTorch-ready UCF-101 baseline; choose Splice if you produce music and want a huge royalty-free sample library with DAW integration.
GeneratedHuman Action Classification vs Storyfile
These tools serve fundamentally different needs and are not direct competitors. Human Action Classification is a free, open-source model for video understanding researchers needing a UCF-101 baseline or low-compute action recognition. StoryFile is a premium, enterprise conversational video AI for creating authentic interactive experiences with real people. Choose Human Action Classification if you are a developer building a video analysis pipeline on a budget; choose StoryFile if you need a polished, cinematically captured digital twin for museum exhibits or legacy preservation.
GeneratedFrequently asked questions
Categories
Topics
Used Human Action Classification? Help shape our editorial sentiment research.