diff --git a/tfjs-backend-cpu/src/backend_cpu.ts b/tfjs-backend-cpu/src/backend_cpu.ts
index 28989e14fba..c4c40f793ab 100644
--- a/tfjs-backend-cpu/src/backend_cpu.ts
+++ b/tfjs-backend-cpu/src/backend_cpu.ts
@@ -21,7 +21,7 @@ import {backend_util, buffer, slice_util, util} from '@tensorflow/tfjs-core';
 import {BackendTimingInfo, DataStorage, DataType, DataValues, KernelBackend, max, NumericDataType, Rank, Scalar, ShapeMap, Tensor, Tensor1D, Tensor2D, Tensor3D, Tensor4D, Tensor5D, TensorBuffer, TypedArray, upcastType} from '@tensorflow/tfjs-core';
 import {kernel_impls} from '@tensorflow/tfjs-core';
 
-const nonMaxSuppressionV3 = kernel_impls.nonMaxSuppressionV3;
+const nonMaxSuppressionV3Impl = kernel_impls.nonMaxSuppressionV3Impl;
 const split = kernel_impls.split;
 const tile = kernel_impls.tile;
 const topkImpl = kernel_impls.topkImpl;
@@ -3117,7 +3117,7 @@ export class MathBackendCPU extends KernelBackend {
 
     const boxesVals = this.readSync(boxes.dataId) as TypedArray;
     const scoresVals = this.readSync(scores.dataId) as TypedArray;
-    return nonMaxSuppressionV3(
+    return nonMaxSuppressionV3Impl(
         boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
   }
 
diff --git a/tfjs-backend-cpu/src/kernels/NonMaxSuppressionV5.ts b/tfjs-backend-cpu/src/kernels/NonMaxSuppressionV5.ts
index 0f982ce55ed..578dccb99f9 100644
--- a/tfjs-backend-cpu/src/kernels/NonMaxSuppressionV5.ts
+++ b/tfjs-backend-cpu/src/kernels/NonMaxSuppressionV5.ts
@@ -18,7 +18,7 @@
 import {NonMaxSuppressionV5, NonMaxSuppressionV5Attrs, NonMaxSuppressionV5Inputs} from '@tensorflow/tfjs-core';
 import {KernelConfig, TypedArray} from '@tensorflow/tfjs-core';
 import {kernel_impls} from '@tensorflow/tfjs-core';
-const nonMaxSuppressionV5 = kernel_impls.nonMaxSuppressionV5;
+const nonMaxSuppressionV5Impl = kernel_impls.nonMaxSuppressionV5Impl;
 import {MathBackendCPU} from '../backend_cpu';
 import {assertNotComplex} from '../cpu_util';
 
@@ -42,7 +42,7 @@ export const nonMaxSuppressionV5Config: KernelConfig = {
     const scoreThresholdVal = scoreThreshold;
     const softNmsSigmaVal = softNmsSigma;
 
-    const {selectedIndices, selectedScores} = nonMaxSuppressionV5(
+    const {selectedIndices, selectedScores} = nonMaxSuppressionV5Impl(
         boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal,
         scoreThresholdVal, softNmsSigmaVal);
 
diff --git a/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV3.ts b/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV3.ts
index c6ec6cc99d1..7015a445b55 100644
--- a/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV3.ts
+++ b/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV3.ts
@@ -39,7 +39,7 @@ export const nonMaxSuppressionV3Config: KernelConfig = {
     const iouThresholdVal = iouThreshold;
     const scoreThresholdVal = scoreThreshold;
 
-    return kernel_impls.nonMaxSuppressionV3(
+    return kernel_impls.nonMaxSuppressionV3Impl(
         boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal,
         scoreThresholdVal);
   }
diff --git a/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV5.ts b/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV5.ts
index 20a2093a752..13c7522d694 100644
--- a/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV5.ts
+++ b/tfjs-backend-webgl/src/kernels/NonMaxSuppressionV5.ts
@@ -17,7 +17,7 @@
 
 import {backend_util, kernel_impls, KernelConfig, TypedArray} from '@tensorflow/tfjs-core';
 import {NonMaxSuppressionV5, NonMaxSuppressionV5Attrs, NonMaxSuppressionV5Inputs} from '@tensorflow/tfjs-core';
-const nonMaxSuppressionV5 = kernel_impls.nonMaxSuppressionV5;
+const nonMaxSuppressionV5Impl = kernel_impls.nonMaxSuppressionV5Impl;
 import {MathBackendWebGL} from '../backend_webgl';
 
 export const nonMaxSuppressionV5Config: KernelConfig = {
@@ -42,7 +42,7 @@ export const nonMaxSuppressionV5Config: KernelConfig = {
     const scoreThresholdVal = scoreThreshold;
     const softNmsSigmaVal = softNmsSigma;
 
-    const {selectedIndices, selectedScores} = nonMaxSuppressionV5(
+    const {selectedIndices, selectedScores} = nonMaxSuppressionV5Impl(
         boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal,
         scoreThresholdVal, softNmsSigmaVal);
 
diff --git a/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV3.ts b/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV3.ts
index 2c4cb226d7e..04866596e2e 100644
--- a/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV3.ts
+++ b/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV3.ts
@@ -15,7 +15,7 @@
  * =============================================================================
  */
 
-import {backend_util, KernelConfig, NonMaxSuppressionV3, NonMaxSuppressionV3Attrs, NonMaxSuppressionV3Inputs, TypedArray} from '@tensorflow/tfjs-core';
+import {kernel_impls, KernelConfig, NonMaxSuppressionV3, NonMaxSuppressionV3Attrs, NonMaxSuppressionV3Inputs, TypedArray} from '@tensorflow/tfjs-core';
 import {WebGPUBackend} from '../backend_webgpu';
 
 export const nonMaxSuppressionV3Config: KernelConfig = {
@@ -34,7 +34,7 @@ export const nonMaxSuppressionV3Config: KernelConfig = {
     const boxesVals = gpuBackend.readSync(boxes.dataId) as TypedArray;
     const scoresVals = gpuBackend.readSync(scores.dataId) as TypedArray;
 
-    return backend_util.nonMaxSuppressionV3(
+    return kernel_impls.nonMaxSuppressionV3Impl(
         boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
   }
 };
diff --git a/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV5.ts b/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV5.ts
index dce6c279dbd..8672d6595ae 100644
--- a/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV5.ts
+++ b/tfjs-backend-webgpu/src/kernels/NonMaxSuppressionV5.ts
@@ -14,7 +14,7 @@
  * limitations under the License.
  * =============================================================================
  */
-import {backend_util, KernelConfig, NonMaxSuppressionV5, NonMaxSuppressionV5Attrs, NonMaxSuppressionV5Inputs} from '@tensorflow/tfjs-core';
+import {kernel_impls, KernelConfig, NonMaxSuppressionV5, NonMaxSuppressionV5Attrs, NonMaxSuppressionV5Inputs} from '@tensorflow/tfjs-core';
 
 import {WebGPUBackend} from '../backend_webgpu';
 export type TypedArray = Float32Array|Int32Array|Uint8Array;
@@ -41,9 +41,10 @@ export const nonMaxSuppressionV5Config: KernelConfig = {
     const scoreThresholdVal = scoreThreshold;
     const softNmsSigmaVal = softNmsSigma;
 
-    const {selectedIndices, selectedScores} = backend_util.nonMaxSuppressionV5(
-        boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal,
-        scoreThresholdVal, softNmsSigmaVal);
+    const {selectedIndices, selectedScores} =
+        kernel_impls.nonMaxSuppressionV5Impl(
+            boxesVals, scoresVals, maxOutputSizeVal, iouThresholdVal,
+            scoreThresholdVal, softNmsSigmaVal);
 
     return [selectedIndices, selectedScores];
   }
diff --git a/tfjs-core/src/backends/backend_util.ts b/tfjs-core/src/backends/backend_util.ts
index 5afeb4bb923..a52e545c3b3 100644
--- a/tfjs-core/src/backends/backend_util.ts
+++ b/tfjs-core/src/backends/backend_util.ts
@@ -31,7 +31,6 @@ export * from '../ops/concat_util';
 export * from '../ops/conv_util';
 export {Activation, FusedConv2DConfig} from '../ops/fused_util';
 export * from '../ops/reduce_util';
-export {nonMaxSuppressionV3, nonMaxSuppressionV5} from './non_max_suppression_impl';
 
 export {BackendValues, TypedArray, upcastType, PixelData} from '../types';
 export {MemoryInfo, TimingInfo} from '../engine';
diff --git a/tfjs-core/src/backends/kernel_impls.ts b/tfjs-core/src/backends/kernel_impls.ts
index fda2a7d1b8f..627eb9b199f 100644
--- a/tfjs-core/src/backends/kernel_impls.ts
+++ b/tfjs-core/src/backends/kernel_impls.ts
@@ -15,7 +15,7 @@
  * =============================================================================
  */
 
-export {nonMaxSuppressionV3, nonMaxSuppressionV5} from './non_max_suppression_impl';
+export {nonMaxSuppressionV3Impl, nonMaxSuppressionV5Impl} from './non_max_suppression_impl';
 export {split} from './split_shared';
 export {tile} from './tile_impl';
 export {topkImpl} from './topk_impl';
diff --git a/tfjs-core/src/backends/non_max_suppression_impl.ts b/tfjs-core/src/backends/non_max_suppression_impl.ts
index caa3c5bb27f..dab7f92caa9 100644
--- a/tfjs-core/src/backends/non_max_suppression_impl.ts
+++ b/tfjs-core/src/backends/non_max_suppression_impl.ts
@@ -32,7 +32,7 @@ interface Candidate {
   suppressBeginIndex: number;
 }
 
-export function nonMaxSuppressionV3(
+export function nonMaxSuppressionV3Impl(
     boxes: TypedArray, scores: TypedArray, maxOutputSize: number,
     iouThreshold: number, scoreThreshold: number): Tensor1D {
   const dummySoftNmsSigma = 0.0;
@@ -45,7 +45,7 @@ export function nonMaxSuppressionV3(
   return result.selectedIndices as Tensor1D;
 }
 
-export function nonMaxSuppressionV5(
+export function nonMaxSuppressionV5Impl(
     boxes: TypedArray, scores: TypedArray, maxOutputSize: number,
     iouThreshold: number, scoreThreshold: number,
     softNmsSigma: number): NamedTensorMap {
@@ -75,7 +75,7 @@ function nonMaxSuppressionImpl_(
 
   for (let i = 0; i < scores.length; i++) {
     if (scores[i] > scoreThreshold) {
-      candidates.push({ score: scores[i], boxIndex: i, suppressBeginIndex: 0 });
+      candidates.push({score: scores[i], boxIndex: i, suppressBeginIndex: 0});
     }
   }
 
diff --git a/tfjs-core/src/ops/image_ops.ts b/tfjs-core/src/ops/image_ops.ts
index f22220898a0..d018355df3e 100644
--- a/tfjs-core/src/ops/image_ops.ts
+++ b/tfjs-core/src/ops/image_ops.ts
@@ -15,150 +15,13 @@
  * =============================================================================
  */
 
-import {nonMaxSuppressionV3, nonMaxSuppressionV5} from '../backends/non_max_suppression_impl';
 import {ENGINE, ForwardFunc} from '../engine';
-import {Tensor, Tensor1D, Tensor2D, Tensor4D} from '../tensor';
-import {NamedTensorMap} from '../tensor_types';
+import {Tensor1D, Tensor2D, Tensor4D} from '../tensor';
 import {convertToTensor} from '../tensor_util_env';
 import {TensorLike} from '../types';
 import * as util from '../util';
-import {nonMaxSuppSanityCheck} from './nonmax_util';
 
 import {op} from './operation';
-export {nonMaxSuppression} from './non_max_suppression';
-
-/**
- * Performs non maximum suppression of bounding boxes based on
- * iou (intersection over union).
- *
- * @param boxes a 2d tensor of shape `[numBoxes, 4]`. Each entry is
- *     `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the corners of
- *     the bounding box.
- * @param scores a 1d tensor providing the box scores of shape `[numBoxes]`.
- * @param maxOutputSize The maximum number of boxes to be selected.
- * @param iouThreshold A float representing the threshold for deciding whether
- *     boxes overlap too much with respect to IOU. Must be between [0, 1].
- *     Defaults to 0.5 (50% box overlap).
- * @param scoreThreshold A threshold for deciding when to remove boxes based
- *     on score. Defaults to -inf, which means any score is accepted.
- * @return A 1D tensor with the selected box indices.
- */
-/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
-/** This is the async version of `nonMaxSuppression` */
-async function nonMaxSuppressionAsync_(
-    boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
-    maxOutputSize: number, iouThreshold = 0.5,
-    scoreThreshold = Number.NEGATIVE_INFINITY): Promise<Tensor1D> {
-  const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppressionAsync');
-  const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppressionAsync');
-
-  const inputs = nonMaxSuppSanityCheck(
-      $boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold);
-  maxOutputSize = inputs.maxOutputSize;
-  iouThreshold = inputs.iouThreshold;
-  scoreThreshold = inputs.scoreThreshold;
-
-  const boxesAndScores = await Promise.all([$boxes.data(), $scores.data()]);
-  const boxesVals = boxesAndScores[0];
-  const scoresVals = boxesAndScores[1];
-
-  const res = nonMaxSuppressionV3(
-      boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
-  if ($boxes !== boxes) {
-    $boxes.dispose();
-  }
-  if ($scores !== scores) {
-    $scores.dispose();
-  }
-  return res;
-}
-
-/**
- * Performs non maximum suppression of bounding boxes based on
- * iou (intersection over union).
- *
- * This op also supports a Soft-NMS mode (c.f.
- * Bodla et al, https://arxiv.org/abs/1704.04503) where boxes reduce the score
- * of other overlapping boxes, therefore favoring different regions of the image
- * with high scores. To enable this Soft-NMS mode, set the `softNmsSigma`
- * parameter to be larger than 0.
- *
- * @param boxes a 2d tensor of shape `[numBoxes, 4]`. Each entry is
- *     `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the corners of
- *     the bounding box.
- * @param scores a 1d tensor providing the box scores of shape `[numBoxes]`.
- * @param maxOutputSize The maximum number of boxes to be selected.
- * @param iouThreshold A float representing the threshold for deciding whether
- *     boxes overlap too much with respect to IOU. Must be between [0, 1].
- *     Defaults to 0.5 (50% box overlap).
- * @param scoreThreshold A threshold for deciding when to remove boxes based
- *     on score. Defaults to -inf, which means any score is accepted.
- * @param softNmsSigma A float representing the sigma parameter for Soft NMS.
- *     When sigma is 0, it falls back to nonMaxSuppression.
- * @return A map with the following properties:
- *     - selectedIndices: A 1D tensor with the selected box indices.
- *     - selectedScores: A 1D tensor with the corresponding scores for each
- *       selected box.
- */
-/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
-function nonMaxSuppressionWithScore_(
-    boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
-    maxOutputSize: number, iouThreshold = 0.5,
-    scoreThreshold = Number.NEGATIVE_INFINITY,
-    softNmsSigma = 0.0): NamedTensorMap {
-  const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppression');
-  const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppression');
-
-  const inputs = nonMaxSuppSanityCheck(
-      $boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold,
-      softNmsSigma);
-  maxOutputSize = inputs.maxOutputSize;
-  iouThreshold = inputs.iouThreshold;
-  scoreThreshold = inputs.scoreThreshold;
-  softNmsSigma = inputs.softNmsSigma;
-
-  const attrs = {maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma};
-
-  const result = ENGINE.runKernel(
-                     'NonMaxSuppressionV5', {boxes: $boxes, scores: $scores},
-                     attrs) as Tensor[];
-
-  return {selectedIndices: result[0], selectedScores: result[1]};
-}
-
-/** This is the async version of `nonMaxSuppressionWithScore` */
-async function nonMaxSuppressionWithScoreAsync_(
-    boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
-    maxOutputSize: number, iouThreshold = 0.5,
-    scoreThreshold = Number.NEGATIVE_INFINITY,
-    softNmsSigma = 0.0): Promise<NamedTensorMap> {
-  const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppressionAsync');
-  const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppressionAsync');
-
-  const inputs = nonMaxSuppSanityCheck(
-      $boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold,
-      softNmsSigma);
-  maxOutputSize = inputs.maxOutputSize;
-  iouThreshold = inputs.iouThreshold;
-  scoreThreshold = inputs.scoreThreshold;
-  softNmsSigma = inputs.softNmsSigma;
-
-  const boxesAndScores = await Promise.all([$boxes.data(), $scores.data()]);
-  const boxesVals = boxesAndScores[0];
-  const scoresVals = boxesAndScores[1];
-
-  const res = nonMaxSuppressionV5(
-      boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold,
-      softNmsSigma);
-
-  if ($boxes !== boxes) {
-    $boxes.dispose();
-  }
-  if ($scores !== scores) {
-    $scores.dispose();
-  }
-  return res;
-}
 
 /**
  * Extracts crops from the input image tensor and resizes them using bilinear
@@ -231,7 +94,4 @@ function cropAndResize_(
   return res;
 }
 
-export const nonMaxSuppressionAsync = nonMaxSuppressionAsync_;
-export const nonMaxSuppressionWithScore = op({nonMaxSuppressionWithScore_});
-export const nonMaxSuppressionWithScoreAsync = nonMaxSuppressionWithScoreAsync_;
 export const cropAndResize = op({cropAndResize_});
diff --git a/tfjs-core/src/ops/non_max_suppresion_with_score_async.ts b/tfjs-core/src/ops/non_max_suppresion_with_score_async.ts
new file mode 100644
index 00000000000..793e873ec22
--- /dev/null
+++ b/tfjs-core/src/ops/non_max_suppresion_with_score_async.ts
@@ -0,0 +1,88 @@
+/**
+ * @license
+ * Copyright 2020 Google Inc. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+import {nonMaxSuppressionV5Impl} from '../backends/non_max_suppression_impl';
+import {Tensor1D, Tensor2D} from '../tensor';
+import {NamedTensorMap} from '../tensor_types';
+import {convertToTensor} from '../tensor_util_env';
+import {TensorLike} from '../types';
+import {nonMaxSuppSanityCheck} from './nonmax_util';
+
+/**
+ * Asynchronously performs non maximum suppression of bounding boxes based on
+ * iou (intersection over union).
+ *
+ * This op also supports a Soft-NMS mode (c.f.
+ * Bodla et al, https://arxiv.org/abs/1704.04503) where boxes reduce the score
+ * of other overlapping boxes, therefore favoring different regions of the image
+ * with high scores. To enable this Soft-NMS mode, set the `softNmsSigma`
+ * parameter to be larger than 0.
+ *
+ * @param boxes a 2d tensor of shape `[numBoxes, 4]`. Each entry is
+ *     `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the corners of
+ *     the bounding box.
+ * @param scores a 1d tensor providing the box scores of shape `[numBoxes]`.
+ * @param maxOutputSize The maximum number of boxes to be selected.
+ * @param iouThreshold A float representing the threshold for deciding whether
+ *     boxes overlap too much with respect to IOU. Must be between [0, 1].
+ *     Defaults to 0.5 (50% box overlap).
+ * @param scoreThreshold A threshold for deciding when to remove boxes based
+ *     on score. Defaults to -inf, which means any score is accepted.
+ * @param softNmsSigma A float representing the sigma parameter for Soft NMS.
+ *     When sigma is 0, it falls back to nonMaxSuppression.
+ * @return A map with the following properties:
+ *     - selectedIndices: A 1D tensor with the selected box indices.
+ *     - selectedScores: A 1D tensor with the corresponding scores for each
+ *       selected box.
+ */
+/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
+async function nonMaxSuppressionWithScoreAsync_(
+    boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
+    maxOutputSize: number, iouThreshold = 0.5,
+    scoreThreshold = Number.NEGATIVE_INFINITY,
+    softNmsSigma = 0.0): Promise<NamedTensorMap> {
+  const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppressionAsync');
+  const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppressionAsync');
+
+  const params = nonMaxSuppSanityCheck(
+      $boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold,
+      softNmsSigma);
+  maxOutputSize = params.maxOutputSize;
+  iouThreshold = params.iouThreshold;
+  scoreThreshold = params.scoreThreshold;
+  softNmsSigma = params.softNmsSigma;
+
+  const boxesAndScores = await Promise.all([$boxes.data(), $scores.data()]);
+  const boxesVals = boxesAndScores[0];
+  const scoresVals = boxesAndScores[1];
+
+  // We call a cpu based impl directly with the typedarray data  here rather
+  // than a kernel because all kernels are synchronous (and thus cannot await
+  // .data()).
+  const res = nonMaxSuppressionV5Impl(
+      boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold,
+      softNmsSigma);
+
+  if ($boxes !== boxes) {
+    $boxes.dispose();
+  }
+  if ($scores !== scores) {
+    $scores.dispose();
+  }
+  return res;
+}
+
+export const nonMaxSuppressionWithScoreAsync = nonMaxSuppressionWithScoreAsync_;
diff --git a/tfjs-core/src/ops/non_max_suppression_async.ts b/tfjs-core/src/ops/non_max_suppression_async.ts
new file mode 100644
index 00000000000..19626a252a6
--- /dev/null
+++ b/tfjs-core/src/ops/non_max_suppression_async.ts
@@ -0,0 +1,73 @@
+/**
+ * @license
+ * Copyright 2020 Google Inc. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+
+import {nonMaxSuppressionV3Impl} from '../backends/non_max_suppression_impl';
+import {Tensor1D, Tensor2D} from '../tensor';
+import {convertToTensor} from '../tensor_util_env';
+import {TensorLike} from '../types';
+import {nonMaxSuppSanityCheck} from './nonmax_util';
+
+/**
+ * Performs non maximum suppression of bounding boxes based on
+ * iou (intersection over union).
+ *
+ * @param boxes a 2d tensor of shape `[numBoxes, 4]`. Each entry is
+ *     `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the corners of
+ *     the bounding box.
+ * @param scores a 1d tensor providing the box scores of shape `[numBoxes]`.
+ * @param maxOutputSize The maximum number of boxes to be selected.
+ * @param iouThreshold A float representing the threshold for deciding whether
+ *     boxes overlap too much with respect to IOU. Must be between [0, 1].
+ *     Defaults to 0.5 (50% box overlap).
+ * @param scoreThreshold A threshold for deciding when to remove boxes based
+ *     on score. Defaults to -inf, which means any score is accepted.
+ * @return A 1D tensor with the selected box indices.
+ */
+/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
+/** This is the async version of `nonMaxSuppression` */
+async function nonMaxSuppressionAsync_(
+    boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
+    maxOutputSize: number, iouThreshold = 0.5,
+    scoreThreshold = Number.NEGATIVE_INFINITY): Promise<Tensor1D> {
+  const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppressionAsync');
+  const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppressionAsync');
+
+  const inputs = nonMaxSuppSanityCheck(
+      $boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold);
+  maxOutputSize = inputs.maxOutputSize;
+  iouThreshold = inputs.iouThreshold;
+  scoreThreshold = inputs.scoreThreshold;
+
+  const boxesAndScores = await Promise.all([$boxes.data(), $scores.data()]);
+  const boxesVals = boxesAndScores[0];
+  const scoresVals = boxesAndScores[1];
+
+  // We call a cpu based impl directly with the typedarray data  here rather
+  // than a kernel because all kernels are synchronous (and thus cannot await
+  // .data()).
+  const res = nonMaxSuppressionV3Impl(
+      boxesVals, scoresVals, maxOutputSize, iouThreshold, scoreThreshold);
+  if ($boxes !== boxes) {
+    $boxes.dispose();
+  }
+  if ($scores !== scores) {
+    $scores.dispose();
+  }
+  return res;
+}
+
+export const nonMaxSuppressionAsync = nonMaxSuppressionAsync_;
diff --git a/tfjs-core/src/ops/non_max_suppression_with_score.ts b/tfjs-core/src/ops/non_max_suppression_with_score.ts
new file mode 100644
index 00000000000..28b945e5d13
--- /dev/null
+++ b/tfjs-core/src/ops/non_max_suppression_with_score.ts
@@ -0,0 +1,84 @@
+/**
+ * @license
+ * Copyright 2020 Google Inc. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+
+import {ENGINE} from '../engine';
+import {NonMaxSuppressionV5, NonMaxSuppressionV5Attrs, NonMaxSuppressionV5Inputs} from '../kernel_names';
+import {NamedAttrMap} from '../kernel_registry';
+import {Tensor, Tensor1D, Tensor2D} from '../tensor';
+import {NamedTensorMap} from '../tensor_types';
+import {convertToTensor} from '../tensor_util_env';
+import {TensorLike} from '../types';
+
+import {nonMaxSuppSanityCheck} from './nonmax_util';
+import {op} from './operation';
+
+/**
+ * Performs non maximum suppression of bounding boxes based on
+ * iou (intersection over union).
+ *
+ * This op also supports a Soft-NMS mode (c.f.
+ * Bodla et al, https://arxiv.org/abs/1704.04503) where boxes reduce the score
+ * of other overlapping boxes, therefore favoring different regions of the image
+ * with high scores. To enable this Soft-NMS mode, set the `softNmsSigma`
+ * parameter to be larger than 0.
+ *
+ * @param boxes a 2d tensor of shape `[numBoxes, 4]`. Each entry is
+ *     `[y1, x1, y2, x2]`, where `(y1, x1)` and `(y2, x2)` are the corners of
+ *     the bounding box.
+ * @param scores a 1d tensor providing the box scores of shape `[numBoxes]`.
+ * @param maxOutputSize The maximum number of boxes to be selected.
+ * @param iouThreshold A float representing the threshold for deciding whether
+ *     boxes overlap too much with respect to IOU. Must be between [0, 1].
+ *     Defaults to 0.5 (50% box overlap).
+ * @param scoreThreshold A threshold for deciding when to remove boxes based
+ *     on score. Defaults to -inf, which means any score is accepted.
+ * @param softNmsSigma A float representing the sigma parameter for Soft NMS.
+ *     When sigma is 0, it falls back to nonMaxSuppression.
+ * @return A map with the following properties:
+ *     - selectedIndices: A 1D tensor with the selected box indices.
+ *     - selectedScores: A 1D tensor with the corresponding scores for each
+ *       selected box.
+ */
+/** @doc {heading: 'Operations', subheading: 'Images', namespace: 'image'} */
+function nonMaxSuppressionWithScore_(
+    boxes: Tensor2D|TensorLike, scores: Tensor1D|TensorLike,
+    maxOutputSize: number, iouThreshold = 0.5,
+    scoreThreshold = Number.NEGATIVE_INFINITY,
+    softNmsSigma = 0.0): NamedTensorMap {
+  const $boxes = convertToTensor(boxes, 'boxes', 'nonMaxSuppression');
+  const $scores = convertToTensor(scores, 'scores', 'nonMaxSuppression');
+
+  const params = nonMaxSuppSanityCheck(
+      $boxes, $scores, maxOutputSize, iouThreshold, scoreThreshold,
+      softNmsSigma);
+  maxOutputSize = params.maxOutputSize;
+  iouThreshold = params.iouThreshold;
+  scoreThreshold = params.scoreThreshold;
+  softNmsSigma = params.softNmsSigma;
+
+  const inputs: NonMaxSuppressionV5Inputs = {boxes: $boxes, scores: $scores};
+  const attrs: NonMaxSuppressionV5Attrs =
+      {maxOutputSize, iouThreshold, scoreThreshold, softNmsSigma};
+
+  const result = ENGINE.runKernel(
+                     NonMaxSuppressionV5, inputs as {} as NamedTensorMap,
+                     attrs as {} as NamedAttrMap) as Tensor[];
+
+  return {selectedIndices: result[0], selectedScores: result[1]};
+}
+
+export const nonMaxSuppressionWithScore = op({nonMaxSuppressionWithScore_});
diff --git a/tfjs-core/src/ops/ops.ts b/tfjs-core/src/ops/ops.ts
index 9797cd5e548..8d88529c810 100644
--- a/tfjs-core/src/ops/ops.ts
+++ b/tfjs-core/src/ops/ops.ts
@@ -124,14 +124,18 @@ export * from './in_top_k';
 
 export {op} from './operation';
 
-// Second level exports.
 import * as losses from './loss_ops';
 import * as linalg from './linalg_ops';
 import * as spectral from './spectral_ops';
 import * as fused from './fused_ops';
 import * as signal from './signal_ops';
 
-import {cropAndResize, nonMaxSuppression, nonMaxSuppressionAsync, nonMaxSuppressionWithScore, nonMaxSuppressionWithScoreAsync} from './image_ops';
+// Image Ops namespace
+import {cropAndResize} from './image_ops';
+import {nonMaxSuppression} from './non_max_suppression';
+import {nonMaxSuppressionAsync} from './non_max_suppression_async';
+import {nonMaxSuppressionWithScore} from './non_max_suppression_with_score';
+import {nonMaxSuppressionWithScoreAsync} from './non_max_suppresion_with_score_async';
 import {resizeBilinear} from './resize_bilinear';
 import {resizeNearestNeighbor} from './resize_nearest_neighbor';
 const image = {
@@ -144,4 +148,5 @@ const image = {
   nonMaxSuppressionWithScoreAsync
 };
 
+// Second level exports.
 export {image, linalg, losses, spectral, fused, signal};