Issues classifying images that are slighly blurred

This may not be considered a bug as the code may be functioning properly so I wanted to drop this here.

System information

• Tested on Ubuntu 22.04 and Windows 10
• TFJS installed via yarn
• TFJS version 3.19.0
• I have written custom code (see snippet below)

Describe the current behavior It seems like TFJS paired with the cocossd model is having issues with images that are blurred or slightly blurry.

Describe the expected behavior Expecting normal image classification as per usual.

Code Snippet For reference, this code is ran in a try/catch so I can manually throw an error if the image arraybuffer request fails. This code is also written in TypeScript which is why I am casting the response from decodeImage to a Tensor3D object for coco to attempt to detect.

import tf, { Tensor3D } from '@tensorflow/tfjs-node'
import cocossd from '@tensorflow-models/coco-ssd'
import axios from 'axios'

// Get Image as arraybuffer
const blob = await axios.get(uri, { responseType: 'arraybuffer' })
    .then((res) => res.data)
    .catch((err) => { throw err; });

// Load Model
const model = await cocossd.load();

// Get Predictions
const predictions = await model.detect(tf.node.decodeImage(blob) as Tensor3D) // Leaving the default boxes and confidence
return predictions;

[ahmedsabie]

@aw1875 Our package is a re-implementation of this tensorflow code, it seems to be an issue with object detection models in general (see this paper)

[aw1875]

@aw1875 Our package is a re-implementation of this tensorflow code, it seems to be an issue with object detection models in general (see this paper)

Alright so it’s just a general issue for image detection as a whole, thank you for the information!