Access the AdShot Benchmark
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
AdShot contains derived frames and audio from third-party video advertisements. Rights to the original ads remain with their owners. Access is granted for non-commercial academic research only.
Log in or Sign Up to review the conditions and access this dataset content.
AdShot Benchmark
Shot-level keyframes, audio, and shot correspondences for 823 pairs of video ads across 194 brands. Each pair links a focal ad to a target ad from the same brand, with a shot-to-shot mapping between them.
| file | size | contents |
|---|---|---|
keyframes_k5_adshot823.h5 |
16.75 GB | 5 RGB keyframes per shot, 819 focal ads |
audio_16k_adshot823.h5 |
1.60 GB | 16 kHz mono audio per shot, 823 focal ads |
adshot823.csv |
443 KB | 823 pairs with shot splits and mappings |
Media is focal-side only — both HDF5 files are keyed by focal_ad_id, and
none of the 671 target ads appear in either. Targets exist in the CSV as
annotation.
Keyframes
/<focal_ad_id> uint8 (n_shots, 5, H, W, 3) lzf
attrs: shot_durations (seconds), relative_path
file attrs: fps=5 (this is k, frames per shot — not frames/second), max_pixels=151200
H, W are fixed within an ad but vary across ads: frames are pre-resized with
Qwen's smart_resize(max_pixels=151200), so stored pixels are what the model
sees. The 5 frames come from the 10th, 25th, 50th, 75th and 90th percentiles of
each shot.
Audio
/<focal_ad_id>/shot_01 ... shot_NN float32 (n_samples,) mono
group attrs: shot_durations, relative_path, sr
file attrs: sr=16000
CSV
One row per pair, 14 columns: brand_id, focal_ad_id, target_ad_id,
mapping, focal_ad_length, focal_shotsplit, target_ad_length,
target_shotsplit, focal_shot_count, target_shot_count, num_mapped_shots,
mapping_degree, focal_relative_path, target_relative_path.
focal_shot_countis the shot count. The keyframe array may hold a couple of extra boundary shots (never fewer), so slicef[ad][:focal_shot_count].mappinghas one entry per focal shot → target shot, orNoneif unmatched:{'shot_01': 'shot_01', 'shot_02': None, ...}. 45% areNone.*_shotsplitlists inclusive end frame indices per shot, so shot 1 is[0, split[0]]and shot 2 is[split[0]+1, split[1]]. Source frame rate is(split[-1] + 1) / ad_length.num_mapped_shotsalways equalstarget_shot_count, andmapping_degreeis always1.0. Neither is informative.
Coverage
Keyframes cover 819 of 823 focal ads. Four source videos have been removed
from YouTube and cannot be rebuilt: dujob0gpAg4, qQLPVjLNGUA,
3yBzrFMnQa0, hET1H6RZVwQ. Audio covers all 823.
Usage
from huggingface_hub import hf_hub_download
import h5py, pandas as pd
repo = "omrastogi/adshot_benchmark"
kf = hf_hub_download(repo, "keyframes_k5_adshot823.h5", repo_type="dataset")
aud = hf_hub_download(repo, "audio_16k_adshot823.h5", repo_type="dataset")
csv = hf_hub_download(repo, "adshot823.csv", repo_type="dataset")
row = pd.read_csv(csv).iloc[0]
ad = row.focal_ad_id
with h5py.File(kf, "r") as f, h5py.File(aud, "r") as a:
if ad in f: # 4 focal ads have no keyframes
shots = f[ad][:int(row.focal_shot_count)] # (n, 5, H, W, 3) uint8
wave = a[ad]["shot_01"][:] # 16 kHz float32
Slice what you need — don't load a 16 GB file into memory.
Provenance
Derived from third-party video advertisements published on YouTube. CC BY-NC 4.0 covers the annotations and packaging; no license is granted to the underlying ads, which remain the property of their brand owners.
- Downloads last month
- 66