Trajectory-pooled conv nets

Closed weights Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences 9.1M parameters May 2015

No estimate

No hardware requirements for this model

The weights for this model have not been published, so it cannot be downloaded or run on your own hardware at any size. It is reachable only through its provider, and no graphics card changes that.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation
Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences
Organisation type
Academia,Academia
Country
Hong Kong, China
Published
19 May 2015
Authors
Limin Wang, Yu Qiao, Xiaoou Tang

What it does

The problem areas the model was built for. A model can carry several of each.

Domain
Video
Task
Action recognition
Numerical format
FP32

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters
9.1M

The input layer takes either a single RGB frame (224x224x3) for the spatial stream or a stack of 10 optical flow frames (224x224x20) for the temporal stream. The first convolutional layer has 96 filters of size 7x7 with stride 2. This is followed by max pooling with size 3x3 and stride 2. The second convolutional layer has 256 filters of size 5x5 with stride 2. After that is another max pooling layer (3x3, stride 2). The third convolutional layer has 512 filters of size 3x3 with stride 1. The fo…

Training data
tokens

They pretrain on ImageNet, and use UCF101 for actions. Its paper says "We introduce UCF101 which is currently the largest dataset of human actions. It consists of 101 action classes, over 13k clips and 27 hours of video data".

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Citations
3,786

Sources

Where this record came from and when it was last checked.

Reference
Action Recognition with Trajectory-Pooled Deep-Convolutional Descriptors
Last updated
28 November 2025

What the numbers mean

Background

Trajectory-pooled conv nets was published by Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences, in Hong Kong, in May 2015. The organisation is categorised as academia,Academia.

It works in Video, and is recorded as doing action recognition.

Because the weights are not available, none of the hardware figures elsewhere on this site apply to it.

Answers

Trajectory-pooled conv nets — common questions

01

Who created Trajectory-pooled conv nets?

Trajectory-pooled conv nets was published by Chinese University of Hong Kong (CUHK),Chinese Academy of Sciences, based in Hong Kong, categorised as academia,Academia.

02

When was Trajectory-pooled conv nets released?

Trajectory-pooled conv nets was published in May 2015. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

03

What is Trajectory-pooled conv nets used for?

Trajectory-pooled conv nets works in Video, and is recorded as handling action recognition. These are the areas it was designed around; they describe intent rather than a hard boundary.

04

What GPU do I need to run Trajectory-pooled conv nets?

None. Trajectory-pooled conv nets is a closed model — its weights were never published, so it cannot be downloaded or run on your own hardware at any price. It is reachable only through its provider.

05

Is Trajectory-pooled conv nets open source?

The licensing for Trajectory-pooled conv nets was never recorded in our source data. We treat unstated licensing as closed, because an unrecorded licence is not one to rely on.

06

How many parameters does Trajectory-pooled conv nets have?

Trajectory-pooled conv nets has 9.1M parameters. The input layer takes either a single RGB frame (224x224x3) for the spatial stream or a stack of 10 optical flow frames (224x224x20) for the temporal stream. The first convolutional layer has 96 filters of size 7x7 with stride 2. This is followed by max pooling with size 3x3 and stride 2. The second convolutional layer has 256 filters of size 5x5 with stride 2. After that is another max pooling layer (3x3, stride 2). The third convolutional layer has 512 filters of size 3x3 with stride 1. The fourth convolutional layer has 512 filters of size 3x3 with stride 1. The fifth convolutional layer has 512 filters of size 3x3 with stride 1. Next is a max pooling layer (3x3, stride 2). The fully-connected layers have 4096, 2048, and 101 neurons respectively. (7*7*20+1)*96 + (5*5*20+1)*256 + (3*3*20+1)*512 + (3*3*20+1)*512 + (3*3*20+1)*512 + 2*4096 + (4096+1)*2048 + (2048+1)*101 = 9106245. That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

Source

Original publication

Record last updated 28 November 2025

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.